I am a final-year Computer Science undergraduate at SRM University AP, working toward inference and ML systems engineering — the layer between a trained model and a system that actually serves it under load.
My work sits across three areas. Applied AI: a deployed multi-agent retrieval platform with causal attribution and a self-recalibrating ranker. Research: symbolic prediction of GPU peak memory for transformer workloads, carried out as undergraduate research. Systems: the CUDA, kernel and serving-path work I am moving into next, because I would rather understand why a model is slow than add another wrapper on top of someone else's API.
I ship production software as well as experiments — full-stack internship work in Node and React, computer-vision pipelines in PyTorch, containerised deployments on Azure — and I care about the engineering discipline around models: reproducibility, evaluation that is not self-congratulatory, and tests that fail for the right reason.
Open to: inference / ML systems engineering roles · applied AI engineering · research internships in efficient ML · open-source collaboration on serving and evaluation infrastructure.
Languages
Frontend
Backend & Databases
ML & Systems
Cloud, DevOps & Tooling
| Domain | Proficiency | Details |
|---|---|---|
| Retrieval & RAG Systems | Advanced | Multi-agent orchestration, hybrid pgvector + BM25 retrieval, leave-one-out causal attribution, self-recalibrating rankers |
| Deep Learning (Vision) | Proficient | PyTorch and OpenCV; real-time crowd monitoring; spatio-temporal attention architectures for human activity recognition |
| Model Evaluation & Reproducibility | Proficient | Paper reproduction under test, regression suites for model behaviour, documenting inconsistencies in published work |
| Interpretable ML | Proficient | VAE-based compression with stacking ensembles, SHAP attribution — published at IEEE ETAACT 2026 |
| GPU Memory & Performance Modelling | Developing | Symbolic prediction of transformer peak memory (UROP research); Nsight-based profiling |
| Inference Engineering | Learning | Paged KV cache, continuous batching, quantisation, CUDA / Triton kernels — active build track |
Nexus — Multi-Agent Research OS
A deployed multi-agent retrieval platform that answers research questions over a document corpus and, critically, explains why it answered that way — each claim is attributed back to the retrieved passages that caused it, using leave-one-out ablation rather than post-hoc citation.
| Stack | Python · PostgreSQL + pgvector · BM25 · Docker · Azure · React |
| Scale | Multi-agent planner / retriever / synthesiser loop over a chunked document corpus |
| Performance | Hybrid dense + lexical retrieval path; self-recalibrating ranker that reweights on observed answer quality |
| Evaluation | Leave-one-out causal attribution; benchmark run on a 20-question self-authored evaluation set |
| Security | Containerised deployment, environment-scoped secrets, no third-party data egress beyond the model provider |
| Impact | Live at nexus-research.me — the reference implementation for how I think retrieval systems should be audited |
| Repository | hashithhh/nexus |
Nexus exists because most RAG systems cannot tell you which retrieved chunk was load-bearing for an answer. Ablating each retrieved passage and measuring the change in the generated answer turns attribution into a measurement instead of a claim. The evaluation set is small and self-authored — stated plainly here rather than dressed up as a benchmark result.
STAD-ConvBi-LSTM — Paper Reproduction Under Test
A from-scratch reproduction of a published spatio-temporal attention architecture for human activity recognition, built as a test-covered codebase rather than a notebook.
| Stack | PyTorch · OpenCV · NumPy · pytest |
| Dataset | UCF11 human action recognition |
| Architecture | Convolutional feature extraction → bidirectional LSTM → spatio-temporal attention |
| Testing | 18 pytest cases covering tensor shapes, attention masking and training-step invariants |
| Finding | 7 internal inconsistencies documented between the published description and a reproducible implementation |
| Repository | hashithhh/stad-convbi-lstm |
The interesting output of this project was not the accuracy number — it was the list of places where the paper could not be reproduced as written. Reproduction is an engineering skill, and writing tests around a research model is how you find out whether you actually understood it.
SYMPEAK — Symbolic GPU Peak-Memory Prediction (UROP · in progress)
Undergraduate research on predicting the peak GPU memory of a transformer workload symbolically — from architecture, sequence length and batch configuration — instead of discovering it by running the job and watching it OOM.
| Stack | Python · PyTorch · CUDA memory instrumentation · symbolic modelling |
| Goal | A parametric envelope for activation + KV-cache + optimiser-state memory across configurations |
| Why | Scheduling and batching decisions in an inference server need a memory answer before the allocation, not after |
| Status | Active undergraduate research (UROP) — results in progress, nothing published yet |
| Repository | Private until results are ready |
This feeds directly into the inference-engine track below: a scheduler that knows the memory envelope of a request can pack the GPU far more aggressively than one that guesses.
Interpretable Wine Quality Prediction (IEEE ETAACT 2026)
| Stack | Python · scikit-learn · PyTorch · SHAP |
| Method | Variational autoencoder compression feeding a stacking ensemble, with SHAP attribution over the learned representation |
| Contribution | Interpretability retained through a compressed latent space rather than sacrificed for accuracy |
| Venue | Published — IEEE ETAACT 2026 |
Four systems projects are in active development. They are listed here for direction, not as finished work — repositories go public when there is something worth reading.
| Project | What It Is | State |
|---|---|---|
| Ledger | Self-maintaining evaluation engine — OpenTelemetry traces scored, human-approved into regression cases, enforced as a CI gate | In progress |
| Crucible | ~100M-parameter language model trained from scratch, with ablations and a scaling study against Chinchilla | In progress |
| Conduit | Inference engine — paged KV cache, continuous batching, quantisation, symbolic memory scheduler | Planned |
| Ignis | CUDA / Triton kernels — staged matmul toward cuBLAS parity, fused softmax and layernorm, FlashAttention-style attention, Nsight roofline analysis | Planned |
Feb 2026 — Apr 2026
Built DeepVision, a real-time crowd monitoring pipeline, as part of the Springboard AI internship programme.
- Implemented detection and tracking over live video streams in PyTorch and OpenCV
- Tuned the inference path for real-time throughput on constrained hardware
- Handled the full loop: data preparation, model integration, evaluation, demo delivery
May 2025 — Jul 2025
Built a resume-to-portfolio application end to end on a Node/Express stack.
- Designed and implemented the parsing pipeline that turns an uploaded resume into structured data
- Built the REST API and the React front end that renders generated portfolios
- Shipped to a live environment and iterated on real user feedback
| Recognition | Details |
|---|---|
| IEEE Publication | Interpretable wine quality prediction — VAE compression, stacking ensemble, SHAP attribution · IEEE ETAACT 2026 |
| Undergraduate Research (UROP) | Selected for undergraduate research on symbolic GPU peak-memory prediction for transformer workloads |
| Production Deployment | Nexus multi-agent research platform, live at nexus-research.me |
| LeetCode 50 Days Badge | 2026 · 315+ problems solved · longest streak 75 days |
| Academic Record | Intermediate MPC 9.1 · SSC 9.8 |
Amazon Web Services
Oracle
NPTEL
Cisco
name: Hashith Alakunta
role: Inference & ML Systems Engineering
location: Andhra Pradesh, India
learning:
- CUDA and Triton kernel programming
- Paged KV cache, continuous batching, quantisation
- Transformer training dynamics and scaling laws
- GPU profiling with Nsight Compute
building:
- Ledger # self-maintaining evaluation engine with a CI gate
- Crucible # ~100M parameter LM from scratch + scaling study
exploring:
- Symbolic prediction of GPU peak memory (UROP research)
- Serving-path bottlenecks in multi-agent retrieval systems
open_to:
- Inference / ML systems engineering roles
- Applied AI engineering roles
- Research internships in efficient ML
- Open-source work on serving and evaluation infrastructure
The model is the easy part. Making it run is the engineering.
