Important

Open to Software Engineer roles (Pune · Philadelphia · remote). Portfolio systems: real-time speech VAD with an F1 gate, offline RAG eval, agents with tool calling and failure handling, and the full stack around them. Benchmark figures link to their reports or CI runs; account statistics link to dated snapshots. Stub metrics are labelled stub.

Fastest contact: [email protected]

---

Start here — three systems, end to end

Project	What it proves	Guardrails	Proof
ai-ml-portfolio — speech VAD + RAG eval + NumPy ML from scratch	Measured VAD F1 gates (real audio + synthetic labels). Offline RAG/LLM eval with retrieval metrics; CI uses stub providers	Acceptance gate fails CI below fixed thresholds; eval falls back to stubs when keys are missing	MODEL_CARD · RESULTS
agent-console · Incident Desk — seeded on-call investigation + protocol console	Labelled, key-free replay of Northwind incidents; planner/tool trace, verdict and seed-based scorecard	Live-capable Nemotron planner via Nebius (not executed); telemetry stays seeded. Original WebSocket reorder/dedupe and RESUME console retained	Agent implementation · Protocol decisions · Pages replay demo
Gravity-SaaS-Agent — multi-tenant agent SaaS starter	MCP-style simulated tool registry, skills scaffold, billing events, operator dashboard	Tier rate limits, Supabase RLS policies, OpenRouter retries/circuit breaker and labelled echo fallback. Channel adapters are simulators, not live Business API	README · CI

---

Engineering metrics — measured, with baselines and a failure case

VAD figures are from the Sep 11, 2026 CI run (batch whole-file processing, excluding warmup). RAG figures are offline stubs. Source audit and verification limits.

Metric	Measured	Gate / baseline	Source
VAD exact F1, noisy — known failure mode	0.7768	≥ 0.75 (clean ≥ 0.92)	challenge_pass.py · MODEL_CARD
VAD soft F1	0.8135	≥ 0.78	same
RAG retrieval (offline stub) — TF-IDF Recall@5 · MRR · nDCG@5	1.000 · 0.981 · 0.986	40 gold QA + 15 adversarial on 10 FAQ docs. BM25-fusion stub Recall@3 0.975. Hallucination 0.000 = stub abstaining	RESULTS.md (STUB, 2026-07-24) · DATA_CARD
Stub cost / latency matrix	Stub p95 ms and estimated $/1k queries from configured prices/token estimates	live path is EVAL_LIVE=1; CI forces EVAL_LIVE=0 (stub)	same RESULTS.md
CI labs + VAD acceptance gate	green	—	Sep 11, 2026

---

Open-source contributions

67 / 75  ██████████████████░░  89.3%

Checked November 14, 2026: 67 company-allowlist merges, within 97 public external merges. Evidence and scope. Latest additions: Comcast contributor-guide links and AWS stdout restoration after printer failures.

By company group (allowlist scoreboard)

Company group	Merged	Focus
Meta	1	Lexical
Amazon/AWS	24	MCP, Stickler, Social Messaging Agent Runtime Toolkit & Visual Asset Management System
Microsoft	10	hve-core, winget-pkgs, physical-ai-toolchain, Semantic Kernel, Playwright docs
Apple	7	Password-manager resources & schema
NVIDIA	5	CUDA Python tests & TensorRT tooling/CI
Adobe	6	React Aria / Spectrum: form validation, CheckboxGroup validation, ComboBox text selection and locale-safe date fields
Sarvam AI	2	Cookbook CI
Google	4	sbsim BUILD cleanup + Gemini cookbook link checker + ADK Go function-tool validation + ADK Go executable code parts
OpenAI	1	codex-security CLI guidance
Netflix	1	DGS codegen Kotlin client generation
Samsung	1	CredSweeper install-guide docs
Moonshot AI	1	kimi-code models.dev fallback
AMD	1	GAIA memory bootstrap confidence parsing
IBM	1	Alchemy Logging: Go goroutine-ID padding
Walmart	1	JSON-to-GraphQL CLI compatibility on Node 22+
Comcast	1	xCOMPASS contributor-guide links

Selected Engineering Impact

Five case studies selected from those 67 merges. Each links to the merged PR and its tests.

Google · ADK-Go: preserve code-execution history across model providers

• Problem: Replaying Gemini conversation history through the OpenAI adapter failed when a turn contained executable code or execution output.
• My contribution: Converted both part types to labeled, fenced text in the Chat Completions and Responses request paths, preserving their order alongside ordinary text and images.
• Validation: Added request-payload regressions for both APIs, including empty code, mixed media and replayed reasoning. Chat Completions tests · Responses tests.
• Outcome: Merged November 6, 2026 · #1956. Request conversion preserves the code and its output as context. This does not execute the code or claim live-provider validation.

Microsoft · Semantic Kernel: enforce URL allowlists after path normalization

• Problem: A raw prefix check let paths such as /api/../admin satisfy an /api allowlist even though the HTTP client requested /admin.
• My contribution: Required both raw and percent-decoded paths to pass the allowlist after dot-segment normalization, preserving empty segments and matching httpx's treatment of terminal dot segments.
• Validation: Added escaping-path and allowed-path cases, comparisons against httpx's wire path, and runner tests proving rejected targets never invoke authentication or send a request. Validator tests · Runner tests.
• Outcome: Merged November 2, 2026 · #14668. The tested traversal and prefix-boundary cases are rejected before outbound work, while valid paths remain accepted.

Amazon/AWS · PostgreSQL MCP: keep database connections on the serving event loop

• Problem: Startup validation initialized PostgreSQL pool state on a temporary event loop, leaving later MCP tool calls with connections bound to a closed loop.
• My contribution: Moved loop-bound startup validation into the MCP server lifespan and added asynchronous cleanup on the same loop, with bounded concurrent connection closes.
• Validation: Regression coverage exercises startup validation followed by the first query, registered lifespan cleanup, startup refusal, and failed or stalled connection closes. Regression tests · Cleanup tests.
• Outcome: Merged November 9, 2026 · #4846. Pool initialization, tool use and shutdown share the serving loop; these tests use controlled database doubles, not a live PostgreSQL deployment.

IBM · Alchemy Logging: align goroutine IDs in Go log headers

• Problem: Pretty-print log headers shifted as goroutine IDs changed digit width, making concurrent logs harder to scan.
• My contribution: Added padding to the widest ID seen so far, using atomic updates because header formatting runs under a read lock. Resetting logger defaults clears the recorded width.
• Validation: Regression tests cover growing and shrinking ID widths, the emitted log header, and reset behavior. Regression tests.
• Outcome: Merged November 11, 2026 · #807. Later headers stay aligned to the widest observed ID; earlier lines are not rewritten.

Walmart · JSON-to-GraphQL CLI: restore Node.js 22 compatibility

• Problem: The CLI's yargs subpath import failed under Node.js 22+, preventing JSON input from being converted into a GraphQL schema.
• My contribution: Switched to the package-root CommonJS export and added a subprocess integration test for the --baseType and --prefix options.
• Validation: The test pipes a JSON fixture through the CLI and checks the generated root and prefixed nested type names. Regression tests.
• Outcome: Merged November 11, 2026 · #50. The CLI uses the compatible yargs entry point while preserving its schema-naming options.

---

Developer tools — install and use

Project	What it does	Install / try	Validation
Codex Insights	Private HTML/JSON workflow reports from Codex sessions; optional shareable aggregate profile. Local-only mode makes zero model calls.	npx @mangeshraut712/[email protected] --local-only	115 tests and 20 synthetic CLI/browser journeys
Mangesh’s Merge Guide	Contribution workflow from issue selection and reproduction through testing, PR review and follow-through.	npx @mangeshraut712/[email protected] install --target codex	Installer, helper and upgrade-safety checks

npm releases verified November 13, 2026. The latest GitHub source includes additional fixes beyond these published versions. For the Insights skill, follow its plugin installation and invoke $insights; Merge Guide installs as $oss-contribution-insights and also supports Claude Code and Cursor destinations. Insights source fixes · Merge Guide source fixes.

---

More projects

• Actions Nurse — Open Agent Hackathon 2026 Wildcard track (author-recorded submission: 28 Oct 2026 4:30 AM IST, hub). CI triage agent: Rules + optional Gemma, Meterless H-MEM, labelled Zetaris fixtures with optional live read-only evidence. submission stamp.

• Asset Intelligence Copilot — FlytBase challenge entry (AI Hackers Collective × Hacktoberfest 2026 Hack Day). Chat copilot for oil-field operators: deterministic tools compute numbers, answers cite record IDs and SOP passages, Gemma 4 for planning/phrasing. Local; MIT.

---

Architecture — patterns across these projects

Conceptual patterns across separate projects; arrows do not assert one integrated deployment.

flowchart LR
  subgraph Client
    B[Browser<br/>Next.js · React]
  end
  subgraph Runtime
    WS[WebSocket agent server<br/>seq reorder · dedupe · RESUME replay]
    API[FastAPI / Express gateway<br/>rate limits · circuit breaker]
  end
  subgraph Models_Tools_Data["Models · Tools · Data"]
    OR[OpenRouter / Workers AI<br/>project-specific routing]
    MCP[MCP-style tool registry<br/>tool calling]
    VAD[WebRTC VAD · speech]
    RAG[BM25 / TF-IDF retrieval<br/>eval gates]
    DB[(Supabase Postgres<br/>RLS policies)]
  end
  subgraph Ops
    CI[GitHub Actions CI<br/>F1 gate · offline eval]
    DEP[Docker · Cloudflare · GitHub Pages]
  end
  B <--> WS
  B --> API
  WS --> MCP
  API --> OR
  API --> RAG
  API --> DB
  B --> VAD
  MCP --> OR
  CI --> DEP

Loading

Pattern	Where it lives	Evidence
Ordered, deduplicated streaming + RESUME	agent-console	reorderBuffer.ts · tests · DECISIONS.md
Rate limits + circuit breaker	Gravity-SaaS-Agent gateway	rate-limiter.ts · circuit-breaker.ts · unit tests
Tenant isolation (RLS)	Gravity-SaaS-Agent	supabase/schema.sql
MCP-style tools (simulated)	Gravity-SaaS-Agent	mcp-client
Workers AI + optional OpenRouter	Hindai	hindai-gemma worker
CI gates, live APIs off	ai-ml-portfolio	ci.yml (EVAL_LIVE: "0", VAD challenge_pass.py)
Containers	agent-console · Gravity-SaaS-Agent	docker-compose.yml · docker-compose.yml

AI engineering stack — tools and dependency evidence

The card lists tools I use or explore. The manifests below verify declared dependencies (checked 2026-11-10), not installed versions or live deployment.

Repo	Declared stack	Manifest
ai-ml-portfolio	NumPy · webrtcvad · librosa · rank-bm25 · scikit-learn	pyproject.toml
agent-console	Next.js 16.4.3 · React 19.1 · TypeScript 5.8 · Turbopack · ws	package.json
Hindai	Next.js 15.5 · React 19.2 · Cloudflare Workers AI	package.json · workers/hindai-gemma
Gravity-SaaS-Agent	Next.js 16.4 dashboard · Express 4 gateway · Supabase	apps
mangeshrautarchive	FastAPI · Uvicorn · Pydantic 2 · httpx	requirements.txt
career-agent-pro	FastAPI · OpenAI SDK · Pydantic 2	backend/requirements.txt
sarvam-ai-cookbook	Sarvam SDK · Flask-SocketIO · websockets	Realtime_Speech_Captioning/requirements.txt

Now — what I'm building in Q4 2026

Track	Where	Latest verified status / evidence
Speech AI / VAD	ai-ml-portfolio · labs/vad	CI + VAD FULL_PASS gate green · Nov 3 run
RAG eval	ai-ml-portfolio · labs/llm-eval	Offline stub report; live mode off in CI
Agent tooling	agent-console · actions-nurse	Incident Desk seeded replay shipped; live Nebius path not executed
MCP apps	Gravity-SaaS-Agent · upstream awslabs/mcp	Registry simulated · 4 awslabs/mcp merges
Developer tools	Codex Insights · Merge Guide	npm releases verified; November 13 source fixes merged; Insights synthetic E2E coverage added
OSS frameworks	scoreboard	59 of 67 allowlist merges landed in Q4

Q4 2026 allowlist merges   Oct  19  ███████████████████░░░░░░░░░
                           Nov  40  ████████████████████████████████████████  (through Nov 14, IST)

OSS merge pace — allowlist merges by month

Merge dates from the scoreboard, grouped by IST month. November is partial (through 2026-11-14).

xychart-beta
  title "Allowlist merges per month (2026)"
  x-axis [Jul, Aug, Sep, Oct, Nov]
  y-axis "Merged PRs" 0 --> 40
  bar [2, 3, 3, 19, 40]

Loading

Repo (top 5)	Merged
awslabs/visual-asset-management-system	12
apple/password-manager-resources	7
microsoft/physical-ai-toolchain	6
adobe/react-spectrum	6
awslabs/stickler	5

In review — open upstream pull requests

View current open upstream PRs on GitHub. Review status changes independently of merge counts; use the live list for current feedback and pending work.

Principles, writing, trajectory

Engineering principles — each one, where it shows up in code

• Measure everything; ship the gate with the number — VAD F1 gate runs in CI and fails below threshold
• Local-first; CI must not depend on a live API — CI sets EVAL_LIVE: "0"; agent-console ships a key-free replay
• Reliability over novelty — tested circuit breaker and rate limiter
• AI should augment, not obscure, the system — Incident Desk shows the planner/tool trace and scorecard (eval.ts)
• Developer experience matters; simple systems scale better — make verify-all runs the tests, VAD gate and offline eval with no API keys

Details: INTERVIEW_NOTES.md

Writing — Technical writings RSS (20 posts in the feed)

Date	Latest posts
Oct 29, 2026	Touchscreen MacBook Pro field notes
Oct 28, 2026	Gemini 4 Argon and Carbon field notes
Sep 29, 2026	OpenAI DevDay 2026: models, ongoing agents, reliable workflows
Sep 10, 2026	Razorpay Vulcan: evaluating a payments foundation model

Trajectory

Figures from the repository's GitHub snapshot, collected 2026-11-13 05:55 IST:

2026 activity	Count
Contribution-calendar total	6,054
Commits	3,700
Pull requests	1,640
Reviews	348
Issues	264

The calendar total and activity-type counts are separate GitHub measures; the rows are not presented as a sum.

Contribution graph · yearly snapshot JSON