Reume

cv.md

Amitava Ghosh

Software Architecture · Platform Engineering · Backend Infrastructure · Workflow Orchestration · Distributed Systems · Reliability & Performance

India amitava.dev@proton.me · (+91) 6294097693 GitHub: ikouchiha47 · Homepage: ikouchiha47.github.io · LinkedIn: segfault-survivor

Profile Summary

Platform engineer with 11+ years of experience, focused on distributed systems, internal developer platforms, and high-throughput backends. Own architecture end-to-end - reliability, performance, multi-team leverage, and cost-aware infra tradeoffs (compute shape, managed data stores, CDN/cache) - from incident RCA to shared SDKs and workflow infrastructure.

Technical Skills

Category	Skills
Languages	Go, Python, JavaScript, Ruby on Rails, Bash, Lua
Data & Databases	PostgreSQL, MySQL, DynamoDB, Redis
Cloud & Platform	AWS, GCP (Vertex), Terraform, CI/CD, Nginx, Datadog, Docker, Linux
Distributed Systems	Event-driven architecture, Workflow orchestration (Prefect), Observability, SDK/DSL Design, Developer Experience
AI / Agent Systems	RAG (BM25, RRF, ColPali), Agent orchestration, Eval frameworks, Langfuse

Professional Experience

Scalarity (Walfin Tech)

Bengaluru, India · Dec 2025 - Feb 2026 Software Architect | Consulting | Python, GCP, embeddings, agent systems

• Owned end-to-end Research, Hypothesis, and Experimentation engines — set problem framing, success metrics, and delivery priority with researchers; the stack takes a falsifiable claim to an executable experiment workspace in ~23 min vs 1–2 weeks of manual literature review and setup, with runs grounded in sandboxed DFT/GPAW compute instead of unverified model output.
• Built and improved search with corrective-RAG and prompt-level reflection (BM25 + vector / RRF, hyp-question embeddings, CHGNet, ColPali, tenant isolation) — nearly doubled papers usefully processed per session (2.1 to 4.7).
• Cut upload drop-off from 34% to 8% and time-to-first-search from ~90s to ~2s via progressive indexing on the workflow engine, streaming results as documents processed.
• Shipped an answer-time rendering layer that turns retrieval and compute outputs (text, figures, tables parsed into queryable datasets, simulation results) into on-the-fly UI components with source citations — so researchers inspect evidence in-place instead of raw model prose.
• Cut tool calls per query from 40 to 22 (~45%) by redesigning the agent strategy and optimizing precision, recall, and memory management, validated with an eval framework against a withheld reference paper.
• Built/deployed models on GCP Vertex with Terraform/IaC and agent-written deploy configs (multi-provider path, including AWS); compressed the materials corpus from ~11.25GB to under 320MB and served formula search from self-hosted SQLite in under 500ms — removing third-party API spend and hard quota ceilings on the critical path.

Affogato (Walfin Tech)

Bengaluru, India · Dec 2024 - Dec 2025 Senior Software Engineer | Python, Go, Postgres, Redis

• Cut media-generation false-fail from ~30% to ~0% with a lean Workflow Engine on existing infra, avoiding months of infra/adoption cost and stopping wrongful refunds/reconciliation risk on future workflow changes.
• Cut per-canvas image transfer from ~4GB to ~10MB with a URL-driven, on-demand transform layer serving each consumer its own optimized variant; upstreamed a memory-usage fix into govips and added canary deployment for the rollout.
• Collapsed prompt-iteration time from hours to minutes by decoupling prompt authorship from backend releases - Markdown units compiled by CI into a versioned artifact store - so backend stopped owning prompt edits.
• Eliminated eve-of-ship integration crunches across a 9-person team by mandating joint API design with mock servers so frontend built against contracts from day one.
• Operated as the de-risking layer for ambiguous initiatives - validated ideas as MVPs, then got leadership buy-in via PRDs/ADRs before scaling to org standards.
• Built the org's agentic-tooling ecosystem (subagents for architecture, implementation, release engineering, bug repro) enforcing eval-driven development scored against code-coverage and quality rubrics, replacing manual review judgment with automated gates.
• Replaced eyeballed video-quality judgment with objective scoring - ffmpeg/VLM analysis of pacing, color, subject tracking, plus sentiment/style evals - the team's first repeatable way to compare AI-generated ads against reference video.

TheBackendCompany

Bengaluru, India · Dec 2023 - Dec 2024 Freelance & Consulting

• Consulted on a bunch of projects from Turing and other sources.
• PM Crash course, built go-batteries (Go backend scaffolding) and a SQLite-WASM CSV tool for large dataset transforms - led to getting noticed and hired.

Sequoia Group

Bengaluru, India · Sep 2020 - Jun 2023 Senior Software Engineer | Go, Redis, AWS, DynamoDB, Python

• Founded the org's Platform Engineering practice by building a company-wide Go SDK - config, logging/PII, HTTP client, notification client, and other cross-cutting modules - that became the mandatory foundation for 6 backend teams and remains in use org-wide.
• Built the org's Notification Service from scratch, consolidating notification-delivery mechanisms that 3 separate projects had built independently, guaranteeing at-least-once and exactly-once delivery through queue-backed, durable writes; still the shared notification path for the org.
• Migrated org-wide file storage to private S3 buckets with signed URLs, closing public-bucket exposure - the custom URL scheme and generic download layer became shared infrastructure other teams built on, extending the same platform foundation.
• Drove a company-wide onboarding SLA change after diagnosing Salesforce-onboarding login outages to MySQL lock contention under repeatable-read isolation, compounded by multi-hour replica lag.
• Fixed a minor-seeming SSO bug whose business impact was outsized - directly credited with retaining 2 client accounts - then updated the shared onboarding docs so every other client company benefited from the same fix.
• Diagnosed a silent CSV-parsing failure to a UTF-8/Windows-1252 encoding mismatch - values displayed correctly in Excel but failed backend validation - then built the org's bulk employee-onboarding pipeline (validation, error reporting, replica-routing proxy) around it.

Gojek

Bengaluru, India · Sep 2016 - July 2020 Product Engineer | Go, Rails, Postgres, Redis, Lua, Kong

• Eliminated a self-inflicted DDoS on the internal auth service (hit by 100-200 microservices) by moving token validation to the edge via Kong API Gateway (~120k RPM peak, ~100M+ bookings/month), migrating 110 APIs across 6 services with scripts that auto-generated Kong configs, establishing Kong as the org-wide standard gateway.
• Restored gateway stability under peak load by diagnosing a retry amplification failure - timeout mismatches, disabled connection reuse, pool contention between Kong and HAProxy - via correlated packet captures and metrics.
• Validated production readiness by replaying ~5x peak traffic in UAT, sizing the deployment before go-live instead of discovering capacity limits in production.
• Replaced manual HAProxy edits and Jira-driven releases with CI-automated, diff-based Kong config deployments, giving security a GitHub PR-based audit trail for API whitelist changes.
• Extended gateway observability (metrics, standardized request logging) beyond reliability - later reused by data engineering to build fraud-detection pipelines.
• Built the Go auth service (Redis-backed sliding-window rate limiting) during the Customer Service platform decomposition, adding location-aware headers and localized error handling that enabled Gojek's later internationalization rollout.
• Built the query-understanding layer for Gojek's global search - a Go service on fastText for query categorization, SymSpell-based typo correction, and query expansion (e.g. matching "M&M" to "M and M") - backed by 3 Elasticsearch indexes kept current via a Kafka pipeline syncing restaurant and location data as listings changed.

Earlier roles

India · Mar 2014 - Sep 2016 Software Engineer | Kreeti, Leftshift | Rails, Node.js, React

• Full-stack product work (social, marketplace, multi-tenant matrimony); Gojek app-review analysis project later acquired with Leftshift engagement.

Open Source / Personal Projects

• streamux (Go) - HTTP/2 library built from spec (RFC 9113 framing/HPACK/stream state machine/flow control, RFC 9218 priorities), validated against h2spec, with pluggable schedulers (DRR, WRR, round-robin) and a C ABI shim for embedding in non-Go runtimes.
• optimux (Go) - self-hosted image/video transform server: on-the-fly resize/transcode by URL, three-tier caching (Nginx edge, tmpfs, S3), dynamic worker scaling that absorbed a 48-request burst with zero drops; 87.9% size reduction on JPEG-to-WebP in ~63ms cold.
• cinestar (Electron/TypeScript) - local-first media search engine indexing personal photos/videos on-device with vision-language embeddings, no cloud upload required.
• Video Eval Framework (WIP) - extends the objective video-scoring work from Affogato (ffmpeg/VLM pacing, color, subject-tracking, sentiment/style evals) into a standalone tool: ffmpeg signal extraction plus LLM judgment against golden video sets for color, mood, storyline, and character consistency.

Education

West Bengal University of Technology

Kolkata, India · 2010 - 2014 B.Tech in Electronics & Communication Engineering

resume.typ

#import "@preview/modern-cv:0.10.0": *

#show link: underline
#show: resume.with(
  author: (
    firstname: "Amitava",
    lastname: "Ghosh",
    email: "amitava.dev@proton.me",
    phone: "(+91) 6294097693",
    github: "ikouchiha47",
    homepage: "https://ikouchiha47.github.io",
    linkedin: "segfault-survivor",
    address: "India",
    positions: (
      "Software Architecture",
      "Platform Engineering",
      "Backend Infrastructure",
      "Workflow Orchestration",
      "Distributed Systems",
      "Reliability & Performance",
    ),
  ),
  date: datetime.today().display(),
  language: "en",
  profile-picture: none,
  colored-headers: true,
  show-footer: false,
)

// Global settings
#set text(size: 9pt)
#set par(leading: 0.55em)

= Profile Summary

Platform engineer with 11+ years of experience, focused on distributed systems, internal developer platforms, and high-throughput backends. Own architecture end-to-end - reliability, performance, multi-team leverage, and cost-aware infra tradeoffs (compute shape, managed data stores, CDN/cache) - from incident RCA to shared SDKs and workflow infrastructure.

#set text(size: 8.5pt)
= Technical Skills
#resume-skill-item(
  "Languages",
  (strong("Go"), strong("Python"), strong("JavaScript"), "Ruby on Rails", "Bash", "Lua"),
)
#resume-skill-item(
  "Data & Databases",
  (strong("Postgresql"), strong("MySQL"), "DynamoDB", strong("Redis")),
)
#resume-skill-item(
  "Cloud & Platform",
  (strong("AWS"), "GCP (Vertex)", "Terraform", "CI/CD", "Nginx", "Datadog", "Docker", "Linux"),
)
#resume-skill-item(
  "Distributed Systems",
  ("Event-driven architecture", "Workflow orchestration (Prefect)", "Observability", "SDK/DSL Design", "Developer Experience"),
)
#resume-skill-item(
  "AI / Agent Systems",
  ("RAG (BM25, RRF, ColPali)", "Agent orchestration", "Eval frameworks", "Langfuse"),
)

= Professional Experience

#resume-entry(
  title: "Scalarity (Walfin Tech)",
  location: "Bengaluru, India",
  date: "Dec 2025 - Feb 2026",
  description: "Software Architect | Consulting | Python, GCP, embeddings, agent systems",
)
#resume-item[
  - Owned end-to-end *Research, Hypothesis, and Experimentation engines* — set problem framing, *success metrics*, and delivery priority with researchers; the stack takes a falsifiable claim to an executable experiment workspace in *\~23 min* vs *1–2 weeks* of manual literature review and setup, with runs grounded in sandboxed DFT/GPAW compute instead of unverified model output.
  - Built and improved search with *corrective-RAG* and *prompt-level reflection* (BM25 + vector / RRF, hyp-question embeddings, *CHGNet*, *ColPali*, tenant isolation) — nearly doubled papers usefully processed per session (*2.1 to 4.7*).
  - Cut upload drop-off from *34% to 8%* and time-to-first-search from *\~90s to \~2s* via progressive indexing on the workflow engine, streaming results as documents processed.
  - Shipped an *answer-time rendering* layer that turns retrieval and compute outputs (text, figures, tables parsed into queryable datasets, simulation results) into on-the-fly UI components with source citations — so researchers inspect evidence in-place instead of raw model prose.
  - Cut tool calls per query from *40 to 22 (\~45%)* by redesigning the agent strategy and optimizing precision, recall, and memory management, validated with an *eval framework* against a withheld reference paper.
  - Built/deployed models on *GCP Vertex* with *Terraform/IaC* and agent-written deploy configs (multi-provider path, including *AWS*); compressed the materials corpus from *\~11.25GB to under 320MB* and served formula search from self-hosted SQLite in *under 500ms* — removing third-party API spend and hard quota ceilings on the critical path.
]

#resume-entry(
  title: "Affogato (Walfin Tech)",
  location: "Bengaluru, India",
  date: "Dec 2024 - Dec 2025",
  description: "Senior Software Engineer | Python, Go, Postgres, Redis",
)
#resume-item[
  - Cut media-generation false-fail from *\~30% to \~0%* with a lean *Workflow Engine* on existing infra, avoiding months of infra/adoption cost and stopping wrongful refunds/reconciliation risk on future workflow changes.
  - Cut per-canvas image transfer from *\~4GB to \~10MB* with a URL-driven, on-demand transform layer serving each consumer its own optimized variant; upstreamed a memory-usage fix into *govips* and added canary deployment for the rollout.
  - Collapsed prompt-iteration time from *hours to minutes* by decoupling prompt authorship from backend releases - Markdown units compiled by CI into a versioned artifact store - so backend stopped owning prompt edits.
  - Eliminated eve-of-ship integration crunches across a *9-person* team by mandating joint API design with mock servers so frontend built against contracts from day one.
  - Operated as the de-risking layer for ambiguous initiatives - validated ideas as MVPs, then got leadership buy-in via PRDs/ADRs before scaling to org standards.
  - Built the org's agentic-tooling ecosystem (subagents for architecture, implementation, release engineering, bug repro) enforcing *eval-driven development* scored against code-coverage and quality rubrics, replacing manual review judgment with automated gates.
  - Replaced eyeballed video-quality judgment with objective scoring - ffmpeg/VLM analysis of pacing, color, subject tracking, plus sentiment/style evals - the team's first repeatable way to compare AI-generated ads against reference video.
]

#resume-entry(
  title: "TheBackendCompany",
  location: "Bengaluru, India",
  date: "Dec 2023 - Dec 2024",
  description: "Freelance & Consulting",
)
#resume-item[
  - Consulted on a bunch of projects from turing and other sources
  - PM Crash course, Built *go-batteries* (Go backend scaffolding) and a SQLite-WASM CSV tool for large dataset transforms - led to getting noticed and hired.
]

#resume-entry(
  title: "Sequoia Group",
  location: "Bengaluru, India",
  date: "Sep 2020 - Jun 2023",
  description: "Senior Software Engineer | Go, Redis, AWS, DynamoDB, Python",
)
#resume-item[
  - Founded the org's *Platform Engineering* practice by building a company-wide *Go SDK* - config, logging/PII, HTTP client, notification client, and other cross-cutting modules - that became the mandatory foundation for *6 backend teams* and remains in use org-wide.
  - Built the org's *Notification Service* from scratch, consolidating notification-delivery mechanisms that 3 separate projects had built independently, guaranteeing at-least-once and exactly-once delivery through queue-backed, durable writes; still the shared notification path for the org.
  - Migrated org-wide file storage to private S3 buckets with signed URLs, closing public-bucket exposure - the custom URL scheme and generic download layer became shared infrastructure other teams built on, extending the same platform foundation.
  - Drove a *company-wide* onboarding SLA change after diagnosing Salesforce-onboarding login outages to MySQL lock contention under repeatable-read isolation, compounded by multi-hour replica lag.
  - Fixed a minor-seeming SSO bug whose business impact was outsized - directly credited with retaining *2 client accounts* - then updated the shared onboarding docs so every other client company benefited from the same fix.
  - Diagnosed a silent CSV-parsing failure to a UTF-8/Windows-1252 encoding mismatch - values displayed correctly in Excel but failed backend validation - then built the org's bulk employee-onboarding pipeline (validation, error reporting, replica-routing proxy) around it.
]

#resume-entry(
  title: "Gojek",
  location: "Bengaluru, India",
  date: "Sep 2016 - July 2020",
  description: "Product Engineer | Go, Rails, Postgres, Redis, Lua, Kong",
)
#resume-item[
  - Eliminated a self-inflicted DDoS on the internal auth service (hit by 100-200 microservices) by moving token validation to the edge via *Kong API Gateway* (\~120k RPM peak, \~100M+ bookings/month), migrating *110 APIs across 6 services* with scripts that auto-generated Kong configs, establishing Kong as the *org-wide standard* gateway.
  - Restored gateway stability under peak load by diagnosing a *retry amplification* failure - timeout mismatches, disabled connection reuse, pool contention between Kong and HAProxy - via correlated packet captures and metrics.
  - Validated production readiness by replaying *\~5x peak* traffic in UAT, sizing the deployment before go-live instead of discovering capacity limits in production.
  - Replaced manual HAProxy edits and Jira-driven releases with CI-automated, *diff-based Kong config* deployments, giving security a GitHub PR-based audit trail for API whitelist changes.
  - Extended gateway observability (metrics, standardized request logging) beyond reliability - later reused by data engineering to build fraud-detection pipelines.
  - Built the Go auth service (Redis-backed sliding-window rate limiting) during the Customer Service platform decomposition, adding location-aware headers and localized error handling that enabled Gojek's later internationalization rollout.
  - Built the query-understanding layer for Gojek's global search - a Go service on fastText for query categorization, SymSpell-based typo correction, and query expansion (e.g. matching "M&M" to "M and M") - backed by 3 Elasticsearch indexes kept current via a Kafka pipeline syncing restaurant and location data as listings changed.
]

#resume-entry(
  title: "Earlier roles",
  location: "India",
  date: "Mar 2014 - Sep 2016",
  description: "Software Engineer | Kreeti, Leftshift | Rails, Node.js, React",
)
#resume-item[
  - Full-stack product work (social, marketplace, multi-tenant matrimony); Gojek app-review analysis project later acquired with Leftshift engagement.
]

= Open Source / Personal Projects
#resume-item[
  - *streamux* (Go) - *HTTP/2* library built from spec (*RFC 9113* framing/HPACK/stream state machine/flow control, *RFC 9218* priorities), validated against *h2spec*, with pluggable schedulers (DRR, WRR, round-robin) and a C ABI shim for embedding in non-Go runtimes.
  - *optimux* (Go) - self-hosted image/video transform server: on-the-fly resize/transcode by URL, three-tier caching (Nginx edge, tmpfs, S3), dynamic worker scaling that absorbed a 48-request burst with *zero drops*; *87.9%* size reduction on JPEG-to-WebP in \~63ms cold.
  - *cinestar* (Electron/TypeScript) - local-first media search engine indexing personal photos/videos on-device with vision-language embeddings, no cloud upload required.
  - *Video Eval Framework* (WIP) - extends the objective video-scoring work from Affogato (ffmpeg/VLM pacing, color, subject-tracking, sentiment/style evals) into a standalone tool: ffmpeg signal extraction plus LLM judgment against golden video sets for color, mood, storyline, and character consistency.
]

= Education
#resume-entry(
  title: "West Bengal University of Technology",
  location: "Kolkata, India",
  date: "2010 - 2014",
  description: "B.Tech in Electronics & Communication Engineering",
)