LLM, RAG & agents
4cooperating agents in AgentDesk
- Multi-agent systems with validated tool calls
- RAG that cites its sources
- Prompt-injection and PII guardrails
- Eval harnesses that catch regressions
Open to AI engineer roles · internships now, full-time 2027 · remote or on-site
AI/ML Engineer · Open-Source Contributor · IEEE Author
I build AI systems that hold up in production, and fix the open-source libraries they run on.
Final-year B.Tech CSE (AI/ML) at GITAM. Three AI internships including PwC, a fine-tuned vision-language model, a multi-agent system, and 50 bug fixes merged across 30 open-source orgs, including unsloth, pydantic, Black, asyncpg, redis-py, Celery and HuggingFace.
Now Fixing edge cases in JAX, dask and pyfakefs found by differential testing against NumPy, SciPy and the real OS.

Counts from GitHub, updated November 2026
Worked at



Code merged into
Most AI demos work in a notebook and fall apart in production. I like the part in between: agents that call tools safely, retrieval that returns the right context, fine-tuned models that are measured honestly, and async code that doesn't freeze an event loop.
That last one turned into CodeHound, a static analyzer I wrote. The bugs it found were fixed and merged in unsloth, agno, pydantic, Black and Weaviate, and I have since had fixes merged in Celery, asyncpg, redis-py, scapy, pyfakefs, pymodbus and Apache Maven. Every one came with a regression test and went through the maintainers' review.
Alongside that I've shipped ML in three internships (PwC, AVAIntern, OneStop AI), had a paper accepted at IEEE ISED 2026, and placed Top 5 at the BITS Pilani AI/ML hackathon.
4cooperating agents in AgentDesk
25→92%extraction validity after QLoRA fine-tuning
−18%inference latency at OneStop AI

PwC · Launchpad Advisory Program
Feb 2026 – Jul 2026Remote

AVAIntern · AvaIntern Edutech
May 2026 – Jun 2026On-site · Visakhapatnam

OneStop AI · 1stop
May 2025 – Aug 2025Remote
Multi-agent customer-support assistant
A planner agent routes each request to knowledge, data and action agents that use schema-validated tool calls. RAG over policy docs with citations, prompt-injection and PII guardrails, human approval for large refunds, and an eval harness that scores every change.
Fine-tuned vision-language model
Qwen2-VL-2B fine-tuned with QLoRA for receipt field extraction and forgery detection. Scored on 120 held-out examples against the zero-shot model, field-extraction validity went from 25% to 92% and forgery F1 from 0.21 to 0.60.
Python static analyzer
AST checks for async-safety and correctness bugs: blocking calls in async code, fire-and-forget tasks, leaked file handles, mutable defaults. Its findings became merged fixes in unsloth, mem0, agno, Black, Weaviate, Xorbits and pydantic-ai. Zero dependencies, CI-ready, with a citable DOI.
AI placement-preparation platform
Generates role-specific study material and mock interviews with RAG over FAISS. A skill-gap scoring layer filters retrieval by the learner's level before prompting, so answers are personal instead of generic, and a roadmap engine builds a study plan from the gaps.
Proxy-proof attendance system
Three checks before attendance counts: face recognition with liveness detection, GPS geofencing within 100 m, and a rotating 6-digit session key. Face detection runs on the device, so biometric data never leaves the phone.
AI mental-health chatbot · Smart India Hackathon 2024
Generic sentence embeddings could not separate neutral text from distress (34% accuracy). Fine-tuning a sentence-transformer on mental-health data with multi-label emotion scoring raised it to 81%, and context tracking cut irrelevant replies by about 60%.
50 pull requests merged across 30 organisations with 400k+ combined stars. Mostly real bug fixes (event-loop blocking, leaked connections, shared mutable state), each with a regression test and reviewed by the project's maintainers.
A time.sleep inside an async route froze the whole event loop for up to 30 seconds. Replaced with a non-blocking wait.
Merged · #6135jax.random.choice dropped the non-sampled dimensions when asked for zero samples, so its output shape disagreed with NumPy's Generator.choice.
Merged · #42430Two grammar-file handles in the pgen converter were opened and never closed.
Merged · #5432A dead isinstance check validated the wrong object and silently discarded the result.
Merged · #13858An exception inside utils.pipeline() leaked a pooled connection on every failure.
Merged · #4345The SQS transport shared queue caches and AWS clients across every connection in a process.
Merged · #2676A missing tcpreplay binary crashed with UnboundLocalError instead of a clear error.
Merged · #5204An exception inside a query_logger() block left the logger attached to the connection, so it kept logging every later query.
Merged · #1373Shutdown returned while a vector-store cleanup task was still running, so process exit cancelled it and left half-deleted batches.
Merged · #6226Every device-identification object shared one class-level dict, so devices overwrote each other.
Merged · #3036Compares ensemble models for predicting student performance and uses SHAP so the drivers behind each prediction can be inspected rather than trusted blindly.
Argues that the gap between reported model compression and real latency or energy gains is a systems problem, and proposes a hardware-aware evaluation protocol.
Uses CodeHound's documented validation record to argue that AI developer tools should be judged on how their rules were validated, not only on whether an output was right.
Combines dual-stream CNNs, YOLO object detection and multi-modal behavioural signals to catch malpractice in both online and in-hall exams.
Surveys deep-learning architectures and deployment strategies for detecting drones and autonomous defence.
BITS Pilani Hyderabad · Techgyan
Placed in the top five at a 24-hour in-person AI/ML hackathon hosted at BITS Pilani, Hyderabad.
NIT Warangal
Ensemble ML with SHAP explainability for student performance prediction, accepted for the IEEE conference proceedings.
Open source
Reviewed and merged by maintainers at HuggingFace, pydantic, Black, redis-py, Celery, unsloth, Apache and more, with code shipping on PyPI.
Government of India
Built SoulSync, an AI mental-health chatbot, for India's largest national hackathon.
LeetCode
91 easy, 108 medium and 13 hard over 127 active days, with an 82-day best streak.
GitHub achievements
Earned for answers accepted by the people who asked them, including one in Snipe-IT that became a merged fix.