skip to content

Austin Dcosta

ai research, then infrastructure.

building things, looking for internships

bengaluru, india (ist)

ai can write code. ai can't have taste.

ludwig van beethoven, by joseph karl stieler, 1820

// about

i liked taking things apart as a kid. remotes, toys, anything that didn't get taken away fast enough. computers were the first thing i could take apart and put back together without getting in trouble.

i study ai/ml at pes university in bengaluru. i started in research and co-authored a paper on discovering llm prompting techniques with knowledge graphs, which was accepted at icann 2026. somewhere along the way i got more curious about what sits underneath: queues, storage engines, processes that crash and come back. that's mostly what i build now.

i also ta data structures & algorithms, do competitive programming, and take quizzing more seriously than i should.


// work

  1. jun – aug 2026 · bengaluru

    summer analyst intern, wealth management technology · morgan stanley

    built an ai-powered validator that checks team names against compliance policy on its own, saving legal and compliance officers weeks of manual review. optimized 10+ sql queries for internal reporting, and shipped backend changes to the team's java codebase.

    learned: how to work properly inside a big collaborative team.

    • java
    • sql
    • llms
  2. aug 2025 – present

    teaching assistant, data structures & algorithms · pes university

    run labs and debugging sessions on trees, graphs, recursion and complexity, and write problem sets.

  3. feb – apr 2025

    project intern, advanced driver assistance · centre for innovation & entrepreneurship, pes university

    real-time vehicle and pedestrian detection with yolov8 (85% accuracy). cut inference latency 30% by switching models based on system load. lane detection and multi-object tracking on embedded hardware.

    learned: how vision models actually behave outside a notebook.

    • pytorch
    • opencv
    • yolov8
    • embedded

// projects

  1. 01mini-vllma from-scratch llm inference engine with a paged kv cache and continuous batching.engine built; no vllm comparison yet2026

    personal project. engine scaffolding, benchmark harness and plots were written by an ai agent.

    • paged kv cache: block pool and block table with slot mapping
    • token-budget scheduler with chunked prefill and recompute preemption
    • prefix cache with lru eviction; blocks are hashed after compute, so same-step requests never read unfinished blocks
    • batched decode over a padded block-table gather, plus an invariant checker
    • token-for-token match with hugging face generate on a tiny random qwen2 (float64, cpu and cuda)
    • profiling showed per-sequence gpu copies dominated decode; each step now uses one host-to-device copy and one batched attention call per layer
    • limits: small model only, no cuda graphs, no fused kernel for the paged path, prefill runs per sequence
    • no comparison against real vllm yet, and no claim that it beats hugging face or static batching
    • python
    • pytorch
    • hugging face
  2. 02distributed job queuea durable single-node job queue in c++ with sqlite storage and lease-based retries.built, not deployed2026

    solo

    // the bottleneck was fsync, not the network

    • multi-threaded socket broker (a thread per connection) with sqlite storage in wal mode
    • at-least-once delivery: workers claim jobs under time-limited leases
    • failed jobs retry with exponential backoff; repeated failures move to a dead-letter queue
    • priority scheduling
    • concurrency safety comes from sqlite's file locking
    • profiling showed per-commit fsync, not networking or threads, as the throughput bottleneck
    • a slow query, diagnosed with explain query plan: a compound index was replaced with two partial indexes
    • a stale-lease race, fixed with a fencing token in the ack/nack protocol
    • c++
    • sqlite
    • posix sockets
  3. 03100 days of cudaa 100-day log of cuda kernels, starting from vector add.in progress2026

    daily practice log

    • planned: vector add and reductions first
    • planned: matmul from naive to coalesced to tiled shared memory, compared against cublas
    • planned: fused softmax with warp reductions
    • planned: a paged-kv gather kernel that reads through a block table
    • no kernel results published yet
    • cuda
  4. 04fault-tolerant ml training runtimea multi-process ml training runtime that survives worker crashes.finished2026

    solo

    • parallel workers across cores that keep training when one dies
    • mmap'd dataset loading
    • crash-safe checkpointing for fast recovery
    • c++
    • posix (fork/exec/waitpid)
    • linux
  5. 05mini-batchera dynamic request batcher in c++, simulated on cpu.built; cpu simulation only2026

    the user wrote the worker in steps 1 and 2. an ai coding agent wrote steps 3 and 4, the benchmarks, plots and docs. the starter skeleton came from a template.

    • pipeline: load generator, request queue, batcher thread, batch queue, then instance threads
    • batches close on a max batch size or a max queue delay
    • simulated model with a fixed cost plus a per-item cost, so it runs without a gpu
    • checks that completed requests equal sent requests and that every callback fires exactly once
    • thread sanitizer runs show no data races
    • c++17
  6. 06bustuba buffer pool manager and b+ tree index from cmu 15-445's bustub.in progress: buffer pool + b+ tree done2026

    cmu 15-445 course project. parts of it were built with claude code sessions.

    • buffer pool manager with page replacement and disk i/o
    • b+ tree index with ordered lookups and range scans
    • c++
  7. 07politicalsima simulated democracy where ai politicians learn to win elections, and voters push back.finished2025

    solo

    • ai candidates trained with reinforcement learning (ppo) shift positions across multiple policy issues to win votes
    • voters live in a small-world social network and influence one another's views
    • shocks like economic crises, pandemics and wars hit the electorate mid-run
    • elections every 20 steps; tracks polarization and voter satisfaction over time
    • python
    • mesa
    • gymnasium
    • stable-baselines3
    • networkx
    • tensorboard

// publications


// writing

writing soon. drafts are in progress.


// offline