Manish K Reddy
city skyline at sunset
city skyline under clouds
hilltop fort among boulders
cloud gate reflecting the skyline
now▸atlanta, gacompiling▸project helix · ontology-first agentsshipping▸numina pilot — first b2b vardrafting▸tpu v6e vs h200 vs mi300x papermerged▸openxla/xla pr #40232reading▸the structure of scientific revolutionslistening▸king krule · alice coltranenow▸atlanta, gacompiling▸project helix · ontology-first agentsshipping▸numina pilot — first b2b vardrafting▸tpu v6e vs h200 vs mi300x papermerged▸openxla/xla pr #40232reading▸the structure of scientific revolutionslistening▸king krule · alice coltrane
Manish K Reddy
atlanta · 2026
§ 01 / about — what i'm pursuing

what i'm pursuing

  1. 01contributing to OpenXLA on the HLO verifier and numerical correctness, with merged PRs to openxla/xla — my favorite kind of contribution is the small one that surfaces five silent crash bugs downstream
  2. 02co-founding Numina AI, where we're building AI-native workflow infrastructure for small CPA firms serving SMBs — currently preparing an angel round and onboarding our first pilot
  3. 03engineering agentic AI infrastructure at CONA Services for Coca-Cola's bottling operations across 500+ servers and 11 bottlers, leading an ontology-first architecture initiative to replace bloated instruction files with a structured knowledge layer
  4. 04writing a cross-platform benchmarking paper comparing TPU v6e, H200, and MI300X — follow-up to my Google Summer of Code work at SDSC / UC Santa Cruz on GPU observability
  5. 05exploring PhD programs in ML systems and compilers — the cascade-bug work and the TPU benchmarking are the threads i'd most want to pull on next
  6. 06finding ways to make a real dent through infrastructure, research, and the occasional well-placed bet
§ 02 / research — compilers, accelerators, agents

OpenXLA — HLO verifier & numerical correctness

i contribute to openxla/xla, focused on the HLO verifier. my most recent merged work, PR #40232 (an async-pair fix), surfaced five silent crash bugs in downstream tests that had been masked for months.

the cascade is what i love about compiler work — one verifier rule, dozens of corrected behaviors. other contributions: a scalar erf saturation fix and test coverage across the HLO surface. recently accepted into the TPU Builder program.

500+
servers patched
across cona ops
5bugs
silent crashes
surfaced by pr #40232
96.6%
throughput recovered
in dr validation
62k–93k
instruction tokens
helix replaces

Cross-platform inference benchmarking

a paper-in-progress comparing TPU v6e, H200, and MI300X across realistic inference workloads — the kind of numbers everyone wants when picking infrastructure but nobody publishes openly.

what i find compelling is how often the intuitive answer turns out to be wrong once you actually measure across batch sizes, sequence lengths, and memory regimes. accelerator choice is downstream of workload shape.

GPU observability for NRP — GSoC 2025

with OSPO at UC Santa Cruz, i built a containerized agentic platform for the National Research Platform (70+ institutions, 3 continents) — ingesting Prometheus metrics to power GenAI narratives and root-cause analyses in the Seam portal.

Project Helix — ontology-first agents at CONA

at CONA Services i lead Helix, replacing bloated agent instruction files (62K–110K tokens) with a structured knowledge layer on Cosmos DB Gremlin, Azure AI Search, and Graphiti. the hypothesis: agent quality degrades as instructions grow, but ontologies don't.

this sits underneath ZENO, a 16-node LangGraph pipeline with seven specialist agents and a Splunk MCP gateway.

Numina AI — workflow infrastructure for small CPA firms

i'm co-founder of Numina, building AI-native workflow infrastructure for the underserved middle of accounting: small CPA firms serving SMBs. what i find compelling is the shape of the market — too small for Big 4, too complex for off-the-shelf SaaS, and workflows that look nearly identical across hundreds of firms.

§ 03 / experience — full-time, summers, fellowships
  1. 01AI Solutions Engineer at CONA Services (Coca-Cola bottling operations), where i shipped the Patching Status Tracking System — React dashboard, Azure Logic Apps orchestration, a 13-tool MCP server on Azure Functions — coordinating maintenance across 500+ servers and 11 teams
  2. 02DR validation system at CONA using FastAPI + React, recovering 96.6% of throughput in disaster scenarios
  3. 03Google Summer of Code 2025 with OSPO at UC Santa Cruz, building GPU observability for the National Research Platform across A100, GH200, L40, and RTX 3090
  4. 04Genisys Venture Analyst at Kaplan Institute, running due diligence on deep-tech projects and helping shape a $6M TechForward investment thesis
  5. 05Grainger Computing Innovation Prize — 2nd Runner-Up for H2.0 Resilience, an explainable-AI flood risk tool that reduced evaluation time by 90%
  6. 06earlier on: machine learning research at VIGA Entertainment on real-time facial motion capture in Unreal Engine, and founding SIGGRAPH BNMIT during undergrad in Bangalore
  7. 07Dartmouth Conrades Distinguished Fellowship, UC Berkeley VC University scholarship, and “Cultural Achiever of the Year” — somewhere between business, research, and being too curious for one lane
build 2026.06v2.0.0eb garamond / jetbrains monohandcoded · atlanta