Sep 2025 – Present
Software Engineer Intern
JavaSpring BootAWSEvent-driven integrationObservability
Now · Software Engineer Intern, ICBC
Harshil Chudasama
Software engineer focused on backend systems, search, and ML infrastructure.
I work on production services, contribute to open-source projects including EF Core, vLLM / GuideLLM and OpenSearch k-NN, and build systems around inference, retrieval, and ranking.
The graph behind this text is a live vector search. Tap it to move the query.
01
Experience
Every role as a span on one timeline, most recent first. Select one to jump to it.
Span
Length
- ICBC1 yr 2 mo, Sep 2025 to Present
- Northeastern Mixed Reality Lab4 mo, Mar 2025 to Jun 2025
- Thermo Fisher Scientific4 mo, May 2024 to Aug 2024
- Northeastern University4 mo, Sep 2024 to Dec 2024
- Felix Payment Systems8 mo, Jan 2023 to Aug 2023
- University of San Francisco4 mo, Sep 2021 to Dec 2021
- Thermo Fisher Scientific2 yr 8 mo, May 2019 to Dec 2021
- St. Joseph's Healthcare Hamilton2 yr, May 2017 to Apr 2019
Mar 2025 – Jun 2025
Northeastern Mixed Reality Lab
AI Research Assistant
PyTorchONNX RuntimeYOLOv8QuantizationvisionOS
May 2024 – Aug 2024
Software Engineer Intern
JavaC++Pythonpybind11SQL
Earlier
- Graduate Teaching AssistantNortheastern University2024
- Product ManagerFelix Payment Systems2023
- Sales Operations AnalystUniversity of San Francisco2021
- Business AnalystThermo Fisher Scientific2019–21
- ResearcherSt. Joseph's Healthcare Hamilton2017–19
02
Open source
A package I maintain, and 16 patches to dotnet, vLLM, OpenSearch, Kubernetes and PyTorch. Merged work sits on the main line, and open work branches off it.
pypi.org/project/adduce
v0.2.0 · MIT
adduce
ML research artifact auditor
Command-line tool that checks whether the results in a machine learning paper can still be traced to the code, data, and settings that produced them. It shows what is missing, file by file, and never claims a project is reproducible.
- 78 checks across 17 categories, from pinned dependencies and random seeds to whether the paper's numbers match the logged results. Each finding points to the file and line and suggests a fix.
- Drafts what reviewers ask for, including NeurIPS and ACL checklist items and an ACM artifact appendix. Labs can add their own rules through a plugin API and run adduce in CI.
Contributions
Aggregates over owned types in GroupBy queries
Queries that summed or averaged an owned-type property inside a GroupBy could not be translated to SQL, so they threw or ran one query per group. They now run as a single SQL query. The fix closed four reported issues, the oldest open since 2022.
GroupBy on nullable columns with NULL keys
Grouping by a nullable column crashed whenever a key was NULL, in every release from 8.0 to 10.0. The fix reads the key correctly and leaves the generated SQL unchanged.
Benchmark support for the Responses API
vLLM's benchmarking tool could not measure the OpenAI Responses API. I added support for it, including correct time-to-first-token for reasoning models.
Reject a cache setting some models silently ignored
For some models a prefix-cache setting was ignored without warning, so a cached request could resume from the wrong state and return different output. vLLM now rejects that setting at startup.
Find the peak load that still meets latency targets
Adds a benchmark mode that raises the load step by step to find the most traffic a model server can handle while still meeting its latency targets. It answers the capacity-planning question in a single run.
Measure requests against latency targets
Lets you set latency targets, such as time to first token, and reports the share of requests that met them. The project's docs already recommended such targets, but the tool had no way to check a run against them.
Confidence intervals in benchmark reports
Reports gave single numbers with no sense of how precise they were. On short runs, the p99 latency was simply the slowest request. Adds a confidence range to the average and each percentile.
Show how long requests wait under heavy load
Under heavy load, the reported latency left out how long a request waited before it was sent at all. I added metrics that make that wait visible.
Exact vector search uses the configured similarity
Exact vector search always scored results by Euclidean distance, whatever similarity the index was set up with, so it could disagree with approximate search on the same index. It now uses the configured similarity.
Vector search in stored percolator queries
Percolation matches new documents against saved queries. Saved vector queries failed on every match, and on multi-shard clusters they returned no results and no error. They now match, scored with exact search.
Accept trace files with extra columns
Trace files with extra columns were rejected before a benchmark started, including the example dataset in the project's own docs. Validation now checks only the columns the tool reads.
Clear error when no model exists yet
Asking for a model before any had been trained returned an error that named an internal index. It now returns a plain "model not found".
Decode list items once
Turning a JSON list into unstructured objects, which the dynamic client and kubectl get -o json both do, decoded every item twice. Reusing the first pass cuts the time by about 40% and the memory by about half, with the same results and errors.
Correct forward-mode gradients for two linear solvers
Forward-mode automatic differentiation returned wrong gradients for two linear-algebra solvers, with no error. Fixes the formulas and closes two issues.
Pooling on very large tensors on Apple GPUs
Pooling layers on Apple GPUs failed on very large tensors, and one of them could write past the end of its output. Adds a path for large tensors and keeps the faster path for normal sizes.
Responses API support in the Rust benchmark tool
vLLM's Rust benchmark tool could not measure the OpenAI Responses API. This ports the Python support I added earlier, and on the same workload both tools report nearly the same throughput and latency.
01
Anytime Inference Planner
Deadline-aware ML serving
ML serving that holds a latency deadline under load by routing each request to a larger or smaller model variant based on how backed up the queue is.
PythonC++ONNX RuntimePyTorchpybind11
2026 · Research
02
Execution Copilot
Multi-venue trading simulator
Simulator for breaking a large trade into smaller orders across venues, comparing classical schedules, contextual bandits, and reinforcement learning.
PythonC++FastAPIPyTorchpybind11
2025 · Research
03
CineMatch
AI assistant and film recommender
Film and TV recommender with an AI assistant. Describe what you are in the mood for, and the assistant searches the catalog and a personal ranking model to answer. It can only recommend titles that exist in the catalog.
GoNext.jsGeminipgvectorPython
2026 · Live
04
Semantic Search Service
Vector-first retrieval with BM25 re-ranking
Document search that matches on meaning rather than exact words, then re-ranks with keyword overlap, metadata, and recency. Relevance is gated in CI.
JavaSpring BootElasticsearchPostgreSQLRedis
2026 · Case study
Also built
- Commerce ServicesEvent-driven e-commerce backend
- BC Cancer Foundation CRMDonor engagement platform
- FaceFit ARBrowser-based virtual try-on
- QuickVerify SDKMobile identity verification
04
Skills
Grouped the way my resume groups them, and ranked within each group by depth.
- Languages
- Python
- Java
- C++17/20
- Go
- TypeScript
- SQL
- Backend
- Spring Boot
- FastAPI
- REST
- gRPC
- Kafka
- Redis
- PostgreSQL
- AWS
- Docker
- ML & search
- PyTorch
- ONNX Runtime
- Elasticsearch
- HNSW
- Embeddings
- Quantization
- Learning-to-rank
- Performance
- pybind11
- Lock-free structures
- JFR
- async-profiler
- perf
- Valgrind
05
Awards
Hackathons and competitions, most recent first.
1st place
2025
Microsoft x Qualcomm On-Device AI Hackathon
Microsoft x Qualcomm
On-device wheelchair navigation on Snapdragon X Elite, fusing YOLOv8 and Whisper at sub-40 ms end-to-end by offloading to the DSP and NPU.
Top 1%
2024
87 of 13,500
IMC Prosperity 2 Trading Competition
IMC Trading
Python market-making and statistical arbitrage: regression fair-value estimation, mean-reversion signals on residuals, and inventory-aware quoting.
Runner-up
2024
Data for Good Hackathon
Telus x City of Toronto
ParkSmart, a real-time parking optimisation proof of concept over Telus mobility data, with heatmaps and routing for downtown Toronto traffic.
3rd place
2024
Climate Resiliency Hack
Northeastern University
Platform helping indigenous communities anticipate and mitigate severe weather events.