Engineer. Builder. Shipper.
AI Developer building scalable, production-grade systems — currently at Delhivery. Previously at InteligenAI, including Spector.AI, where I built MCP-orchestrated agent frameworks, document intelligence pipelines, and vision systems across AWS, Azure, and GCP.
I work across NLP, computer vision, and ML — multi-agent reasoning with LangGraph, RAG with hybrid retrieval, document classification over 500K+ records at 99% accuracy, and vision models with YOLO, SAM, and Gemini. B.Tech ECE'24, HBTU Kanpur · Gurugram, India.
Experiences
Software Engineer
Delhivery
- Building the internal agentic AI platform — the runtime, gateway, and governance layer other teams' agents run on.
- Architected a deterministic graph-based workflow engine (LLM, agent, branching, sandboxed-code, guardrail, and concurrent-iteration nodes) with staged versioning and zero-redeploy rollback.
- Drove adoption of a multi-provider LLM gateway across services — unblocking Claude-backed agents by root-causing provider-specific routing.
- Built per-agent cost governance with cache-aware pricing and hard budget short-circuits.
- Instrumented the platform with Langfuse distributed tracing and a Presto/Trino outcome-evaluation pipeline.
- Cut agent tool-calling iterations 8→3 via SOP-scoped lazy tool loading over six internal MCP servers.
- Built MCP-orchestrated AI agent frameworks with LangGraph for multi-step reasoning in industrial plants.
- Delivered document intelligence (99% classification, 90% extraction on 500K+ docs) with a fine-tuned custom NER for field extraction.
- Built RAG chatbots with hybrid search, RRF, and HyDE, plus a natural-language-to-SQL/Mongo agent querying 20+ tables.
- Engineered async backends with FastAPI and Celery, deployed behind Apache load balancing.
- Delivered vision systems with YOLO, SAM, GroundingDINO, and Gemini for industrial defect detection.
- Fine-tuned SDXL (DreamBooth) for subject-consistent image generation.
- Deployed on AWS, Azure, and GCP with Docker, Kubernetes, and CI/CD.
Electronics & Design Engineer
Eleqzee Motors
- Internship at IIT Kanpur — electronics design and prototyping for motor systems.
jun – jul 2023
IoT Engineer
Invariance Automation
- Internship at IIT Kanpur — IoT systems and automation for industrial applications.
jan – feb 2023
Education
B.Tech, Electronics & Communication Engineering
Harcourt Butler Technical University
Kanpur, India · Graduated 2024
2020 – 2024 ↗
Intermediate (CBSE), PCM
Little Flower House
Kanpur, India
2019
Projects
● sep 2026
JED Red-Team Attack
- Algorithmic red-teaming submission for the JED agent-security benchmark (Kaggle, hosted by OpenAI, Google & IEEE), stress-testing tool-using AI agents.
- Searches multi-step attack paths that carry an agent from untrusted input to an unsafe action.
- Reverse-engineered the benchmark's guardrail from its SDK to find structural bypasses rather than clever prompts.
- Iterated across real leaderboard submissions to a best score of 89/1000.
Agent SecurityLLM Red-TeamingPythonGPT-OSSGemma
Groot — Plant Disease Assistant
- Multimodal, multilingual plant-disease assistant (Kaggle Gemma hackathon) — photo, voice note, or text in 7 languages.
- Returns cited, research-grounded diagnoses, running fully local.
- A ReAct agent with an information-confirmation gate that refuses to diagnose until both plant and symptoms are confirmed.
- Grounded in a ChromaDB vector store.
GemmaFastAPIChromaDBRAGWhisperReAct
Variational Autoencoder from Scratch
- From-scratch PyTorch implementation of autoencoders and VAEs — deterministic AE → MLP-VAE → CNN-VAE.
- Trained across MNIST, Fashion-MNIST, and CIFAR-10.
- Implements the reparameterization trick and Gaussian ELBO by hand.
- Written notes deriving the math from the original paper.
PyTorchVAEDeep LearningGenerative Models
Object Semantic Separator
- Pairs YOLOv8 detection with Meta's SAM for pixel-precise object separation.
- YOLO's bounding boxes become the region prompts that focus SAM into clean per-object masks, then removes the background.
- Turns coarse detection into precise segmentation without training a segmentation model from scratch.
- Evaluated with IoU/Dice overlap.
YOLOv8SAMOpenCVPython
Source Code Converter
- A source-to-source code translator on Streamlit + Google Gemini.
- Upload a file, name a function, pick a target language — it converts that function.
- Self-scores the conversion quality on a 1–5 scale with a second Gemini pass.
- Custom extension-based language detection handles raw files that off-the-shelf libraries misclassify.
Generative AILLMGeminiStreamlitPython
Image Caption Generator
- Preprocessed Flickr8K with data cleaning and VGG-16 image feature extraction.
- Integrated captions and features with LSTM models optimized for effective learning.
- Evaluated caption accuracy with BLEU scores against human references.
Deep LearningPythonNLPComputer VisionLSTMVGG-16
Skills
JunJulAugSepOct
NovDecJanFebMar
AprMay
Contact
Interested in a conversation? Reach me here:
- Emailsrivastavashashank46@gmail.com
- Phone+91 84295 48955
- LinkedInlinkedin.com/in/shashank-srivastava ↗
- RésuméDownload ↗