AI / LLM ENGINEER HYDERABAD, IN 4 YRS PHENOM PEOPLE

Aditya
Dasika.

Building production AI systems — from LLM fine-tuning to physics-informed neural networks.

I architect agentic conversational pipelines, fine-tune LLMs at scale, and implement PINNs from scratch in PyTorch. Mechanical engineering by training, self-taught into production ML.

akhil.dasika47@gmail.com adityadasika21 linkedin download résumé (.pdf)

97%

NLU accuracy

5.72 rps

prod throughput

98.6%

success rate

10k+

monthly txns shipped

01 / About

I'm an AI/LLM Engineer at PHENOM — architecting 12-node LangGraph agentic pipelines, fine-tuning LLMs on AWS, and deploying to vLLM in production.

On the side, I implement physics-informed neural networks from scratch in PyTorch — heat equations, Burgers', Schrödinger.

Mechanical engineering by training, self-taught into production ML. I think about AI systems differently because I came in through physics, not through web apps.

02 / Work

Things I've shipped to production.

Where the model meets the load test. Two roles, two different kinds of "shipping": one full-stack solo build, one agentic ML system at enterprise scale.

Phenom People — AI / LLM Engineer

chatbotCXAgent · 12-node LangGraph pipeline

Architecting an enterprise conversational AI: a 12-node LangGraph pipeline with discrete owners for Concierge, Intent Classifier, Dialogue Manager, NLG, CTC, and seven more. QLoRA fine-tuned Qwen3.5-2B on 4,569 production records, deployed to vLLM on AWS EC2 g5.xlarge.

Built the eval framework end-to-end — pass@k, GPT-4.1-mini judge rubrics across all 12 nodes — and lifted intent accuracy from 87% to 97% by replacing a generic intent model with a multi-task fine-tuned ModernBERT (intent + NER, joint heads).

87 → 97%

NLU accuracy

multi-task ModernBERT · intent + BIO NER

5.72RPS

load tested

QLoRA Qwen3.5-2B · vLLM · g5.xlarge

98.6%

success rate

at peak RPS, mixed-traffic profile

7 → 78%

test coverage

unit + node-level + LLM-judge eval

LangGraph LangChain FastAPI Redis PostgreSQL FAISS FalkorDB ModernBERT Qwen3.5-2B vLLM PyTorch Langfuse Kafka Unsloth QLoRA / NF4 AWS EC2 g5.xlarge

Production debugging — three I'm proud of

// parallel tool calls in agent loop
spurious parameter generation when nodes ran concurrent tool calls — isolated invocation context per branch, fixed.
// JSON schema
$ref recursive loops crashing the pipeline on certain payloads — flattened schemas at boundary, fixed.
// vLLM + Qwen Mamba+attention hybrid
undocumented at the time: required merged weights and --enforce-eager on vLLM — found it, fixed it, deployed.

Blueleaves Farms — Full-Stack Engineer

POS system · 5 locations · 2 brands · React + Firebase

Sole developer, end-to-end. Designed and built a multi-location point-of-sale from scratch — billing & checkout, inventory with SKU-grouping (variants collapsed into single SKUs), cash handling, Razorpay payments, thermal receipt printing, role-based accounts for 20 staff, and a React Native mobile POS variant.

Owned the schema, the React + Material UI web app, the Expo mobile app, the Python Cloud Functions backend, and the on-site network and hardware setup at all 5 locations. Daily and monthly sales reporting across two brand catalogs. 3-hour solo SLA — no escalation path.

10k+

monthly transactions

2 yrs in production

200+

daily peak orders

concurrent across locations

city locations

multi-location stock sync

staff accounts

role-based access · cashier / admin

Scope I owned, top to bottom

// product
billing & checkout · SKU-grouped inventory · cash handling · Razorpay payments · thermal printer receipts · role-based auth (20 staff)
// reach
web POS (React + MUI) · mobile POS (React Native + Expo) · multi-location stock across 5 cities · 2 brand catalogs (30 SKUs)
// platform
schema design · Firestore + Auth + Hosting · Python Cloud Functions on GCP · on-site network & hardware setup at every location
// solo on-call
3-hour SLA. No platform team, no escalation path. 2 years live, still running.

React Material UI React Native Expo Firebase Firestore Firebase Auth Cloud Functions (Python) GCP Razorpay SDK thermal printer SDK Photoshop Blender on-site sysadmin

03 / Selected projects

Built from scratch, mostly without frameworks.

Hover any card to see the equation or core idea underneath.

P / 01

Physics-Informed Neural Networks

PyTorch · from scratch · 3 PDEs

Implemented automatic differentiation for physics residuals — no PDE-specific frameworks. Three classic equations, one codebase. Two-phase optimization (Adam → L-BFGS), complex-valued networks where the physics demands it.

PyTorch autograd L-BFGS complex-valued nets

~3%

heat eq. error

Raissi
parity

burgers' benchmark

8.3e-5

schrödinger loss

∂u/∂t = α · ∂²u/∂x²

P / 02

Karpathy GPT — from scratch

transformer · autograd · 5k steps

Full transformer following "Let's Build GPT". The point wasn't to ship — it was to internalize: Q/K/V projections, q @ k.T, the actual VJP graph through create_graph, LayerNorm, teacher forcing. Now I read papers differently.

PyTorch multi-head attention VJPs teacher forcing

4.39 → 1.86

loss curve

5,000

training steps

tiny
shake

corpus

attn(Q,K,V) = softmax(QKᵀ/√d)V

P / 03

mmBERT — multi-task NLP

ModernBERT · 2 task heads · uncertainty weighting

ModernBERT backbone, two heads: 4-class intent + BIO NER. Uncertainty-weighted loss (Kendall et al.). Diagnosed validation loss divergence — fixed with differential learning rates, stratified splits, label smoothing. Shipped as a candidate for the chatbotCXAgent intent pipeline.

ModernBERT multi-task HuggingFace PEFT stratified splits

intent
+ NER

heads

σ-weighted

loss

prod
candidate

handoff

L = Σᵢ (1/2σᵢ²) Lᵢ + log σᵢ

P / 04

IndexNotes — paper knowledge graph

Neo4j · Ollama · PDF → graph pipeline

Reading a paper isn't enough — I wanted to see how ideas connect. Neo4j-backed concept graph: ~40 concepts, 17+ relationships across three papers (Uni-HamGNN, PINN Part I & II). PDF in, structured graph out.

Neo4j Ollama PDF parsing graph schema

~40

concepts

17+

relationships

papers indexed

G = (V_concepts, E_relations)

04 / Skills

What I reach for, by cluster.

Not a checklist of buzzwords — what I actually use, organized by where in the system it lives.

LLM & GenAI

LangGraph LangChain RAG Graph RAG QLoRA fine-tuning vLLM LLM deployment prompt engineering Langfuse multi-agent systems

ML & Deep Learning

PyTorch Transformers BERT / ModernBERT PINNs multi-task learning uncertainty weighting FAISS vector search

Infra & Backend

FastAPI Redis PostgreSQL FalkorDB Neo4j Kafka AWS EC2 SageMaker Docker

Frontend & Tools

React Unsloth HuggingFace PEFT Ollama Git

05 / Contact

Let's talk.
I'm open to roles.

Email · best

akhil.dasika47@gmail.com

GitHub

@adityadasika21

/in/adityadasika21