Skip to content

Abhishek Srivastava

  • Junior ML Engineer at Arivihan Technologies
  • B.Tech, IIT Kanpur (Minor in ML & Applications)
See the workDownload resume

> I build AI systems for the problems that emerge when prototypes meet real users, real data, and real-world constraints—scale, reliability, latency, cost, retrieval, and observability....

About

Building production AI at the intersection of intelligence and infrastructure

I care about the engineering around intelligence — retrieval over messy data, agent orchestration, model routing, latency, cost, failure recovery, observability, and whether an answer can be traced back to its source.

At Arivihan, I build production AI systems: a multi-agent tutoring stack with six specialist agents, 30+ tools, RAG, long-term memory, multimodal capabilities, and a fault-tolerant serving layer designed for large-scale traffic.

Before that, I built Graph RAG and enterprise knowledge systems at InfraDock, and retrieval infrastructure and 150+ enterprise integrations at UnifyApps.

My earlier work was in computer vision and robotics — from defect inspection on a steel-plant production line to terrain perception for a legged robot at IIT Kanpur.

Now
Junior ML Engineer, Arivihan
Focus
Agentic AI · GenAI · Computer Vision · Machine Learning · System Architecture
Based
Noida, India
Studied
B.Tech (Mechanical), IIT Kanpur

Experience

Where the systems shipped

    • Built a production multi-agent LLM tutor with 6 agents, 30+ tools and RAG, running on GPT-5.4/5.4-mini and Gemini 3.1 Flash.
    • Engineered hybrid agent memory — summarization plus embedding retrieval — for persistent context across a user session.
    • Cut token costs 75% by migrating inference from GPT-5.4 to GPT-5.6-Luna without losing answer quality.
    • Built a fault-tolerant AI serving pipeline on FastAPI, Redis, DynamoDB and SQS/DLQ supporting 100K concurrent users.
    • Reduced live-chat analytics from 1,500 to 1 DynamoDB query per class, lifting AI-category coverage from 4% to 100%.
    • Shipped a complaint and feedback agent using FastAPI live retrieval and RAG over OpenAI Vector Stores for query resolution.
    • Multi-agent LLM
    • RAG
    • FastAPI
    • DynamoDB
    • Redis
    • SQS/DLQ

    • Built parsing and ingestion pipelines for construction DPRs, Work Orders and BOQs, holding 92% field accuracy.
    • Modelled entities into a work-graph for Graph RAG, improving multi-hop query accuracy 35% over vector-only retrieval.
    • Added layout-aware chunking, dedup and incremental re-indexing, cutting ingestion per project pack to under 10 minutes.
    • Revamped the analyst frontend with conversational querying and citation drill-down to the source BOQ row.
    • Graph RAG
    • Document AI
    • Ingestion
    • Knowledge Graphs

    • Engineered scalable RAG retrieval services and ingestion pipelines for enterprise PDF, image and Excel sources.
    • Designed RBAC-enabled agentic indexers with semantic filters, cutting query latency 25% to sub-200ms.
    • Shipped 150+ API integrations in Java and Groovy, and evaluated OCR agents for 15% higher extraction accuracy.
    • Enterprise RAG
    • RBAC
    • Java
    • Groovy
    • OCR

Selected work

Four systems, and what each one had to survive

Problem

Generating tutoring questions with a single prompt is unpredictable: the model silently skips steps, and nothing about the run is inspectable or testable.

Approach

A LangGraph planner emits a structured JSON execution plan per request, then vision, solver and evaluator agents run against typed Pydantic state so every hop is deterministic and can be asserted on in tests.

Decisions

  • Planner-first over reactive chaining — the plan is data, so it can be logged, diffed and replayed.
  • Pydantic typed state at every node, which turns agent bugs into schema failures instead of silent bad output.
  • A separate evaluator agent so quality is scored by a different context than the one that produced the answer.

Pipeline

  1. planneremits JSON execution plan
  2. visionparses diagrams and figures
  3. solverderives the worked solution
  4. evaluatorscores and gates the output
specialised agents
4
typed execution plans
JSON

Stack

  • LangGraph
  • Python
  • Pydantic
  • LLM Evals

Problem

A legged robot that trusts its camera equally on every frame walks into terrain it cannot handle — glare, motion blur and low light all degrade the traversability estimate without warning.

Approach

A vision-proprioception fusion framework: a ResNet-50 reliability estimator scores each frame on brightness and blur, GaNAV predicts terrain traversability, and a ROS2 DWA planner weighs the two in real time.

Decisions

  • Score image reliability explicitly rather than assuming every frame is usable — the planner leans on proprioception when vision is weak.
  • Trained GaNAV on 8,000+ RUGD images to cover outdoor terrain classes the lab actually operates in.
  • Deployed as ROS2 nodes so the planner runs on-robot at control-loop rate.

Pipeline

  1. cameraRGB frames from the Go2
  2. reliabilityResNet-50 brightness + blur score
  3. GaNAVterrain traversability segmentation
  4. DWA plannerROS2 terrain-aware local planning
more robust navigation
25%
RUGD images trained on
8,000+

Stack

  • PyTorch
  • ResNet-50
  • GaNAV
  • ROS2
  • OpenCV
  • Unitree Go2

Problem

Pellet quality was judged by eye, so defective batches were caught late and sent back for manual reprocessing.

Approach

Annotated 3,000+ images in CVAT and Fiji, trained YOLOv8 detection and U-Net segmentation models in PyTorch, and deployed the pipeline on-site for real-time inspection.

Decisions

  • Detection and segmentation together — YOLOv8 locates pellets, U-Net measures them.
  • Built the annotation set in-house so the label taxonomy matched the plant's own defect vocabulary.
  • Shipped to the line rather than to a notebook, which is where the reprocessing saving came from.
inspection accuracy
92%
less manual reprocessing
25%

Stack

  • YOLOv8
  • U-Net
  • PyTorch
  • OpenCV
  • CVAT

Problem

Rotor-bearing vibration signals captured in LabVIEW were being read by hand, which is slow and inconsistent across operators.

Approach

Redesigned a VGG-16 CNN in TensorFlow over the LabVIEW signal set for multi-class fault classification, used t-SNE to check that the fault classes were genuinely separable, and shipped a Tkinter GUI for operator-side inference.

Decisions

  • t-SNE separability analysis before trusting the accuracy number.
  • A desktop GUI so the model is usable by the lab without a Python environment.
multi-class accuracy
95%
separability verified
t-SNE

Stack

  • TensorFlow
  • VGG-16
  • t-SNE
  • Tkinter
  • LabVIEW

Toolkit

What I reach for

Languages
  • Python
  • C++
  • Java
  • SQL
  • MATLAB
  • TypeScript
ML / DL
  • PyTorch
  • TensorFlow
  • Hugging Face
  • Scikit Learn
  • OpenCV
  • XGBoost
  • LightGBM
GenAI & RAG
  • LangChain
  • LangGraph
  • FAISS
  • Pinecone
  • GraphRAG
  • LLM Evals & Guardrails
  • Agentic Workflows
  • Hugging Face
  • Transformers
Cloud & MLOps
  • AWS Lambda
  • API Gateway
  • DynamoDB
  • S3
  • EC2
  • Vertex AI
  • Docker
  • Kubernetes
  • Jenkins
  • GitHub Actions
Backend & Data
  • FastAPI
  • Flask
  • Streamlit
  • REST APIs
  • PostgreSQL
  • MySQL
  • MongoDB
  • Git
  • Linux

Education

Where it started

Indian Institute of Technology Kanpur

2022 — 2026

B.Tech, Mechanical Engineering

Minor in Machine Learning & Applications

Relevant coursework

  • Introduction to Machine Learning
  • Statistical Natural Language Processing
  • Data Structures & Algorithms

Entrance exams

AIR 2889
JEE Advanced 2022among 1.6 lakhs+ candidates
AIR 2306
JEE Mains 2022among 1 million+ candidates
AIR 122
NEST 2022among 22,000+ candidates
AIR 387
IISER Aptitude Test 2022among 22,000+ candidates