Building reliableAI infrastructurefor production systems.
AI/ML engineer building vision-sensing systems, multi-agent platforms, and production infrastructure — currently leading Optisense AI and contributing fixes upstream to NVIDIA and OpenTelemetry.
Professional Snapshot
Selected Work
Production-grade systems across multi-agent orchestration, analytics, and core infrastructure.
Full-Stack Web · AI Content Pipeline
IPL Verse
Full-Stack Daily IPL Gaming Platform
Problem: IPL fans needed an engaging daily gaming platform based on real historical data, but parsing raw datasets into playable trivia required extensive curation.
Build: Designed end-to-end with Next.js App Router and Supabase, featuring real-time multiplayer Arena battles. Engineered an AI pipeline via Groq API to generate and deduplicate 3,600+ trivia questions.
ShippedA live public product with an active user base, supporting authentication, leaderboards, and multiple daily game modes.
Multi-Agent AI · Code Generation
API Forge AI
Agentic SDK Generation Platform
Problem: Converting OpenAPI specs into production-ready SDKs manually is error-prone and time-consuming, requiring constant updates as APIs evolve.
Build: Built a multi-agent LangGraph platform featuring Planner, Validator, Diagnoser, Coder, and Executor agents with self-healing code correction.
ShippedAn automated pipeline that ingests OpenAPI specs and outputs fully functional SDKs with Dockerized backend services.
AI Automation · Full-Stack
AgencyOS AI
AI-Powered Influencer Agency Workflow
Problem: Influencer marketing agencies handle dozens of inbound brand inquiries daily, requiring hours of manual brief extraction, creator matching, and CRM logging.
Build: Built a fully autonomous, multi-agent n8n system integrating Telegram bots, Groq LLaMA 3.3, and Supabase to orchestrate lead intake, deal intelligence, and creator matching.
ShippedA production-ready automated pipeline that matches creators to brand briefs in under 5 minutes with a Next.js real-time monitoring dashboard.
Computer Vision · ML Research
Autonomous Driving VQA
Fine-tuning BLIP on driving scenes with semantic evaluation.
Problem: BLIP performs well on general VQA but struggles in specialized domains like road scenes.
Build: Fine-tuned BLIP on BDD100K driving imagery with procedurally generated QA pairs.
Shipped96.00% lexical accuracy — but 83.47% true BERTScore.
Core Infrastructure
Optimizing runtimes, correcting memory leaks, and building reliability in widely-used production frameworks.
fix(auth): isolate OAuth2 client per flow in console/MCP auth handler
The Context
“ConsoleAuthenticationFlowHandler tracked concurrent OAuth2 flows per-flow, but stored the OAuth2 client used to redeem the authorization code in a single shared attribute instead of per-flow state — so a second flow starting before an earlier one's browser login finished would silently overwrite the first flow's client.”
The Impact
Added a client field to the per-flow state, populated right after constructing the OAuth2 client, and switched the redirect callback to read from that isolated state instead of the shared attribute — with a regression test that reproduces the concurrent-flow corruption pre-fix and passes post-fix.
fix(finetune): apply CLI config overrides in finetuning runtime
The Context
“The --override flag in the nat finetune CLI command was correctly parsed but previously ignored by the runtime.”
The Impact
Explicitly merged user-provided overrides into the config dictionary, producing a fully prepared run configuration.
fix memory leak lru cache
The Context
“Using functools.lru_cache on class instance methods created permanent strong references, preventing garbage collection across 25 critical core utilities.”
The Impact
Replaced improper lru_cache usage with cached_property and manual dictionary caches, fixing 25 memory leaks and preventing runaway memory usage.
fix(core): resolve lru_cache memory retention in DiscoveryMetadata
The Context
“Bare @lru_cache decorators on DiscoveryMetadata's static methods pinned the 128 most-recently-seen dynamic Pydantic classes in memory indefinitely, quietly extending their lifetime in long-running processes like nat mcp serve.”
The Impact
Replaced the leaking cache with a WeakKeyDictionary-based strategy so the garbage collector reclaims dynamically generated classes automatically, and resolved lingering typing/docstring feedback from a related prior PR in the same pass.
Professional Trajectory
Experience across deep learning research, distributed AI systems architecture, and early-stage startup engineering.
December 2025 – Present
Optisense AI
Project Lead, AI Vision Systems
MeitY GENESIS Grant Project, IIT Bhilai
- ·Selected for the competitive MeitY GENESIS EiR program at IIT Bhilai; awarded a ₹10 lakh government grant to prototype AI-driven vision sensing systems for smart infrastructure.
- ·Lead end-to-end development of the prototype — architecture, backend system design, AI inference pipelines, and Dockerized deployment — building low-latency vision intelligence modules for constrained edge and cloud hardware.
May 2025 – January 2026
Visual Intelligence and Learning Lab, IIT Hyderabad
AI/ML Research Intern
Research Internship
- ·Reproduced and benchmarked DaViT and GenAD models on real-world sensor datasets, documenting generalization behavior across distribution shifts; implemented 3D Gaussian Splatting for neural rendering and dynamic object removal under deployment-constrained latency and memory budgets.
- ·Built a modular BLIP-based Visual Question Answering pipeline on BDD100K with custom annotation tooling and evaluation harnesses, enabling rapid iteration across model configurations.
Leadership & Community
Engagement across engineering clubs, hackathon planning, developer groups, and peer mentoring.
Team Member — AEGIS
Inter IIT Tech Meet 14.0
Architected a multi-agent trading platform via a 7-stage pipeline with SHAP/LIME explainability.
AI/ML Lead
Google Developer Groups, IIT Bhilai
Directed deep learning workshops and authored widely adopted ML resources.
Core Member
Data Science & AI Club
Organized intensive technical bootcamps and mentored 30+ engineering peers.
Student Mentor
IIT Bhilai
Guided incoming junior cohorts on advanced coursework strategy.
Lessons from shipped fixes.
Short technical takeaways pulled from real, merged pull requests and research — not things I read, things I had to work out.
NVIDIA/NeMo-Agent-Toolkit
“Bare @lru_cache on static methods doesn't leak in the traditional sense, but it pins the 128 most-recently-seen dynamic classes in memory indefinitely — and linters that catch instance-method leaks don't flag static methods. A module-level WeakKeyDictionary keeps the caching benefit while letting the GC reclaim classes that are no longer referenced.”
open-telemetry/opentelemetry-python
“Overriding __setitem__ blocks direct assignment, but CPython's C-level dict methods — update, pop, clear — bypass Python's method overrides entirely and mutate in place anyway. Real immutability means explicitly overriding every mutating method, not just the obvious one.”
API Forge AI — Agentic SDK Generation Platform
“Early versions had one agent generate and fix its own SDK output — correction quality was inconsistent, because the model that made a mistake usually couldn't see it as a mistake. Splitting the pipeline into distinct Planner, Validator, Diagnoser, Coder, and Executor agents put a fresh evaluation pass between generation and correction, so errors got caught instead of rationalized away.”
Autonomous Driving VQA — Research at IIT Hyderabad
“Fine-tuning BLIP on BDD100K driving scenes hit 96% lexical accuracy but only 83.47% true BERTScore — the model was matching surface tokens, not meaning. It reframed evaluation for the rest of the project: exact-match metrics overstate correctness on open-ended VQA, and semantic similarity scoring is the one that actually matters.”
Honors & Metrics
Key software metrics and honors received from national programs, academic incubations, and developer hackathons.
0
Other Experiences
0
Awards Won
0
Roles & Programs
0
OSS Contributions
Awards & Honors
3rd Place
AI Hackathon (Pathway)
Architected a real-time financial analytics streaming platform.
2nd Place
TechSprint (GDG)
Built a Flutter-based, Gemini-integrated campus application.
Selected Participant
MeitY GENESIS EiR Program, IIT Bhilai
Awarded a ₹10 lakh government grant to prototype AI-driven vision sensing systems for smart infrastructure.