Skip to content
View IamHetPatel's full-sized avatar
♾️
Hustling
♾️
Hustling

Highlights

  • Pro

Block or report IamHetPatel

Block user

Prevent this user from interacting with your repositories and sending you notifications. Learn more about blocking users.

You must be logged in to block users.

Content in all repositories owned by your account will be closed.
Maximum 250 characters. Please don’t include any personal information such as legal names or email addresses. Markdown is supported. This note will only be visible to you.
Report abuse

Contact GitHub support about this user’s behavior. Learn more about reporting abuse.

Report abuse
IamHetPatel/README.md

Het Patel

AI systems engineer in training. M.Sc. Applied AI at TU Dresden (avg 1.5), working student at a Dresden AI startup. I build the layer around LLM and ML systems that decides whether you can trust them: retrieval, structured extraction, evaluation gates, tracing, and data pipelines that survive a handover.

Open to a 30-ECTS master's thesis in industry (late 2026 / early 2027) in AI systems or ML infrastructure preferrably.

Works I've done

EchoClaim · voice claims-intake agent with grounded retrieval (Big Berlin Hack 2.0) Python · FastAPI · LiveKit/Twilio · Gemini 2.5 Flash · GLiNER2 · Qdrant · GitHub Actions.

  • Fine-tuned GLiNER2 chosen over Gemini for extraction on latency (69 ms / 0.61 F1 vs 2,729 ms / 0.86 F1).
  • Retrieval MRR 0.314 → 0.805 over BM25 on a 32-query adversarial golden set.
  • CI eval gate with tolerance bands, LLM-judge groundedness 0.703, content-hashed prompt registry, PII redaction.
  • Every number reproduces via docs/EVAL.md.

BINNVC · structural-prior 3D EM segmentation (TU Dresden research project, team of 3) PyTorch · MONAI · SLURM on the Capella cluster

  • A differentiable adjacency-prior loss for a 3D U-Net cut floating-membrane violations from 0.41 to 0.03 with no Dice regression.
  • Component-level evaluation traced the real failure to fragmented topology; caught a duplicate-test-crop bug that was double-counting scores.

PatchCore anomaly detection · training-free industrial defect localisation

  • Frozen WideResNet-50-2, greedy coreset to ~1 % of patches, FastAPI heatmap endpoint, CPU-only.
  • 0.984–1.000 image AUROC on MVTec bottle, carpet and leather.

DataWhisper-AI · plain-English questions → MongoDB aggregation pipelines, each stage streamed over SSE (START Hack, St. Gallen)

API Gateway · JWT auth, Redis rate limiting, response cache, Prometheus + Grafana · Retail Pulse · Go + RabbitMQ async image processing

Writing

Elsewhere

Springer ACSS 2024 co-author (undergraduate NLP work) · 6+ hackathons, Deploy on Scroll award at EthIndia 2023 · 250+ LeetCode (Java)

hetp943@gmail.com · LinkedIn

Pinned Loading

  1. EchoClaim EchoClaim Public

    VOICE THAT UNDERSTANDS · CLAIMS THAT MOVE; We built a voice agent that already knows the caller.

    Python 1

  2. API-Gateway API-Gateway Public

    An API Gateway with built-in rate limiting, request/response caching, and load balancing using Node.js, Redis (for caching), NGINX/HAProxy (for load balancing), Express, MongoDB.

    JavaScript 1

  3. retail-pulse retail-pulse Public

    A scalable backend service for managing and processing store visit images using Go, Docker and asynchronous processing with RabbitMQ.

    Go

  4. DataWhisper-AI DataWhisper-AI Public

    TypeScript 1

  5. mvtec-anomaly-detection-model mvtec-anomaly-detection-model Public

    Python