Aahan
Bagga

ML engineer at Pronto Networks and CS + Data Science at UC Berkeley (’27). I’ve built the AI workloads; now I want to build the chips under them.

Actively searching for Summer 2027 internships in AI/ML, hardware, and software engineering.

Run the NPU Email me

pick a block — or tab into it

Interactive floorplan of a generic system-on-chip die A reference drawing of an SoC die showing a four-core CPU cluster, an NPU systolic array, shared L3 cache banks, a GPU shader array, on-chip SRAM, media and I/O, and the memory PHY, joined by a network-on-chip. Select any block to read what it does. SOC-REF · illustrative floorplan · not to scale CPU ×4 NPU L3 CACHE SRAM GPU MEDIA + I/O MEM PHY
SoC reference floorplan seven blocks · hover or select one

A generic die layout, drawn to explain the parts rather than to copy any real product. The network-on-chip is what everything else negotiates with.

SHT 02

How an NPU actually multiplies

This grid is how AI chips do math. Each square holds one number. Data flows in from the left, each square multiplies and adds, and the answers flow out the bottom. Press play.

Array size

More rows amortise the fill and the drain.

Cycle
0 / 61
MACs issued
0
Cells busy now
0 / 256
Array utilisation
0%
Pass efficiency
—

fill and drain cost K + N − 2 cycles on every pass

weight held, idle multiply-accumulating wavefront activation in partial sum out
SHT 03

Experience

newest at the top

  1. Machine Learning Engineer

    Full-time

    Pronto Networks · Walnut Creek, CA · May 2026 – Present

    • Built a graph neural network anomaly-detection pipeline, raising recall from 40% to 93% and halving false positives.
    • Designed root-cause isolation feeding LLM-driven remediation.

  2. Software Consultant

    Contract

    IBM · Berkeley, CA · Jan 2026 – May 2026

    • Built an autonomous TORCS driver on IBM Granite, cutting lap times from 60+s to 28s.

  3. Machine Learning Researcher

    HERO2 · Berkeley, CA · Jan 2026 – May 2026

    • Built a multimodal route engine (HA*, UCS) and cut Neo4j query time by 96%.

  4. Data Science Intern

    CerebrAIx Technologies · San Francisco, CA · Feb 2025 – Sep 2025

    • Built Xpredict, a multi-LLM + SBERT pipeline scoring candidate-job fit, benchmarked across 5+ models.

  5. Director of Data Science

    Kids Teach Tech · San Francisco, CA · Sep 2020 – Dec 2024

    • Co-founded a free AI/data science nonprofit; grew it to 100+ volunteers and a 300+ student camp at UC Berkeley.
SHT 04

Projects

HGNN Anomaly Detection & RCA

Site-level heterogeneous graph models feeding an org-wide aggregator, paired with a two-stage root-cause pipeline that names the failing device and drafts the remediation steps.

  • PyTorch
  • GNNs
  • LLM agents
  • Anomaly detection
40% → 93% recall

Multimodal Navigation Engine

A routing engine combining HA* and Dijkstra’s UCS to optimise across time, cost, and energy, layered over live transit APIs. Neo4j query time fell from 50s+ to under two seconds.

  • Graph algorithms
  • Neo4j
  • Transit APIs
  • Python
96% faster queries

Autonomous TORCS Driver

A real-time autonomous racing agent on IBM Granite for IBM’s AI Racing League, tuning steer gain, target speed, and braking logic live against vehicle telemetry.

  • Real-time control
  • IBM Granite
  • Telemetry
  • Python
60s → 28s lap

Xpredict

A multi-model inference pipeline blending GPT-4, Claude, and Gemini with SBERT embeddings to score candidate-job fit, benchmarked across quality, cost, and latency.

  • LLM orchestration
  • SBERT
  • REST API
3-model ensemble

AgentLedger

A B2B proxy enforcing real-time cost budgets, kill switches, and anomaly detection across production AI-agent workflows, built after a competitive gap analysis in agent observability.

  • Node.js
  • FastAPI
  • Redis
  • Docker
28 companies analysed Source

foreman

An agentic software factory chaining Port, Bright Data’s Scraper Studio, and SigNoz, with a physical Reachy Mini robot standing in as the human-approval gate before any change ships.

  • Agentic AI
  • Observability
  • Python
hardware in the loop Source

Physical SOC

A security agent that finds new CVEs on the live web, patches and tests them in a sandbox, and opens a PR. Then a Reachy Mini robot reads the CVE out loud and nothing merges until a person approves that exact commit by voice.

  • Agentic AI
  • MCP
  • Robotics
  • Python
4-condition merge gate Source

Hollow

A test-quality checker. Coverage only proves a line ran, so Hollow deletes each function’s body and reruns its tests. If they still pass, nothing was actually checking that function.

  • Python
  • pytest
  • Testing
  • CLI
finds untested code Source

W2CAI

A full-stack AI product-discovery platform for Asian fashion — parallel scraping and LLM extraction pipelines feeding automated sourcing with integrated reviews.

  • Next.js
  • React
  • LLM extraction
7k+ entries Source
SHT 05

Toolbox

Languages
  • Python
  • C
  • Java
  • SQL
  • Go
  • R
ML, AI & data
  • PyTorch
  • TensorFlow
  • Scikit-Learn
  • Pandas
  • NumPy
  • LangChain
  • Transformers
  • RAG
  • Agentic AI
  • A/B testing
  • EDA
Systems & backend
  • Docker
  • Git
  • REST APIs
  • FastAPI
  • Express.js
  • Redis
  • MLOps
Learning now
  • Computer architecture
  • Verilog / RTL
  • CUDA
  • Memory hierarchy
  • Roofline analysis
  • Dataflow scheduling

the bottom row is where my evenings go

SHT 06

Education

University of California, Berkeley

Class of 2027

B.S. Computer Science & B.S. Data Science — early admit, top 1.1% of applicants · GPA 3.6 / 4.0

  • Data Structures & Algorithms
  • Machine Learning
  • Agents & Generative AI
  • Database Systems
  • SICP
  • Artificial Intelligence
  • Data Science Principles
  • Computer Security
  • Linear Algebra
  • Multivariable Calculus
Copied to clipboard