Toronto · York ’28

Max
Naidenov.

Hi, I’m Max. I study computer science at York University and I’m an ML / AI Engineering Intern at SCHWAI. This site has my projects, my work history and my writing.

GitHub activity

GitHub · naidx0

  1. 3 on 2026-07-19
  2. 28 on 2026-07-20
  3. 4 on 2026-07-21
  4. 0 on 2026-07-22
  5. 1 on 2026-07-23
  6. 79 on 2026-07-24
  7. 15 on 2026-07-25
  8. 0 on 2026-07-26
  9. 0 on 2026-07-27
  10. 1 on 2026-07-28
  11. 7 on 2026-07-29
  12. 113 on 2026-07-30
  13. 0 on 2026-07-31
  14. 0 on 2026-08-01
  15. 0 on 2026-08-02
  16. 50 on 2026-08-03
  17. 52 on 2026-08-04
  18. 64 on 2026-08-05
  19. 29 on 2026-08-06
  20. 90 on 2026-08-07
  21. 16 on 2026-08-08
  22. 16 on 2026-08-09
  23. 14 on 2026-08-10
  24. 45 on 2026-08-11
  25. 113 on 2026-08-12
  26. 84 on 2026-08-13
  27. 1 on 2026-08-14
  28. 43 on 2026-08-15
  29. 10 on 2026-08-16
  30. 58 on 2026-08-17
  31. 64 on 2026-08-18
  32. 52 on 2026-08-19
  33. 36 on 2026-08-20
  34. 30 on 2026-08-21
  35. 97 on 2026-08-22
  36. 24 on 2026-08-23
  37. 37 on 2026-08-24
  38. 45 on 2026-08-25
  39. 132 on 2026-08-26
  40. 94 on 2026-08-27
  41. 114 on 2026-08-28
  42. 45 on 2026-08-29
  43. 8 on 2026-08-30
  44. 23 on 2026-08-31
  45. 21 on 2026-09-01
  46. 123 on 2026-09-02
  47. 12 on 2026-09-03
  48. 125 on 2026-09-04
  49. 544 on 2026-09-05
  50. 573 on 2026-09-06
  51. 439 on 2026-09-07
  52. 65 on 2026-09-08
  53. 285 on 2026-09-09
  54. 692 on 2026-09-10
  55. 165 on 2026-09-11
  56. 39 on 2026-09-12
  57. 45 on 2026-09-13
  58. 69 on 2026-09-14
  59. 36 on 2026-09-15
  60. 21 on 2026-09-16
  61. 14 on 2026-09-17
  62. 84 on 2026-09-18
  63. 24 on 2026-09-19
  64. 12 on 2026-09-20
  65. 28 on 2026-09-21
  66. 193 on 2026-09-22
  67. 156 on 2026-09-23
  68. 560 on 2026-09-24
  69. 223 on 2026-09-25
  70. 30 on 2026-09-26
  71. 7 on 2026-09-27
  72. 0 on 2026-09-28
  73. 0 on 2026-09-29
  74. 28 on 2026-09-30
  75. 41 on 2026-10-01
  76. 107 on 2026-10-02
  77. 16 on 2026-10-03
  78. 5 on 2026-10-04
  79. 144 on 2026-10-05
  80. 315 on 2026-10-06
  81. 75 on 2026-10-07
  82. 1 on 2026-10-08
  83. 314 on 2026-10-09
  84. 297 on 2026-10-10
  85. 172 on 2026-10-11
  86. 325 on 2026-10-12
  87. 4 on 2026-10-13
  88. 82 on 2026-10-14
  89. 0 on 2026-10-15
  90. 0 on 2026-10-16
  91. 0 on 2026-10-17
  92. 0 on 2026-10-18
  93. 0 on 2026-10-19
  94. 0 on 2026-10-20
  95. 0 on 2026-10-21
  96. 0 on 2026-10-22
  97. 0 on 2026-10-23
  98. 0 on 2026-10-24
  99. 0 on 2026-10-25
  100. 0 on 2026-10-26
  101. 0 on 2026-10-27
  102. 0 on 2026-10-28
  103. 0 on 2026-10-29
  104. 0 on 2026-10-30
  105. 0 on 2026-10-31

· public calendar, no tokenlessmore

Work in private repositories is not in this chart.

Featured posts

Selected projects

The other 3 are on the projects page.

selvage

live

Team knowledge ends up scattered across Slack, wikis, code and databases, and I wanted one place to search all of it that did not ship my employer's data to anybody. So the embeddings run on your own CPU and the whole thing is MIT-licensed.

Selvage is a thin, hybrid-searchable layer over the places team knowledge actually lives, Slack, wikis, code, and databases. Connectors normalize every source into one Postgres + pgvector table, so the same query serves a person at a CLI, an agent over MCP, or the web UI. Local-first and MIT-licensed, with embeddings running on your own CPU.

your datanormalizestorerank (no LLM)fuseclientsfolder / gitSlackany Postgrescustomconnectorconnectorsshared schemaPostgres+ pgvectorfull-textGINvectorHNSWrare-termIDFRRF k=60recency decaycross-encoderCLIMCP serverweb UI
Selvage retrieval: connectors normalize every source into one Postgres + pgvector table; three ranked lists (full-text, vector, rare-term) fuse with RRF, decay by recency, and are reranked by a local cross-encoder before reaching the CLI, MCP server, or web UI.

sequence

live

I kept watching architecture diagrams drift away from the code they described. This one reads the code back and fails the build the moment they disagree, down to a wrong HTTP method.

Purpose. Coding agents describe a system confidently whether or not they know it. Sequence reads the code first, so every answer can point at the file and line behind it.

your codescanmapaskanswerrepositoryon diskscannerimports, calls, routesthe mapevery edge citedyour question+ slice of the mapanswer with filespath drawnMCP toolsfor other agents
Sequence: it reads your repository with static analysis, builds a map where every connection points at the file and line that proves it, and answers your question from the relevant slice of that map, with the files it used and the path drawn.

ml-harness

in progress

Training on one consumer GPU taught me the two questions nobody asks out loud: is this data worth the electricity, and is the number at the end real. This is my attempt to answer both before the run starts.

Most people who think they need to fine-tune a model don't. A clearer prompt, a few examples, or a way to look things up usually does the job, and training is the slow, expensive fix that fails quietly when it is the wrong one. The good training tools (Unsloth, Hugging Face's peft and trl) all start after you have decided to train. ML Harness is the step before that.

askmeasuredecideactcheckyour ask,plain wordslook atthe datascore themodel you havefive gatescheaper fixor a scripttrain witha recipescore againstthe baseline
ML Harness: your ask goes to a local engine, which measures your data and the model you have, then checks five gates. A cheaper fix is done instead of training, a missing piece is named, and training runs only when all five pass, and is scored against the baseline after.

schwai

live

My internship project at SCHWAI. A knowledge base that gets better the more the company uses it, with open-source models fine-tuned on the company's own writing rather than a generic one bolted on the side.

Project Brain is SCHWAI's internal enterprise knowledge base, built to improve itself through feedback loops from live usage, open-source models fine-tuned on the organization's own knowledge so the company operates at higher efficiency over time. It's one app in a larger multi-app platform I build across schwai.com on FastAPI, React, Supabase, Redis, and Cloudflare Stream.

company contentingeststoreretrieveask in plain englishbooks workereverythingelse producedbackgroundworkersPostgres+ SupabaseRedisfast pathRAG over thecompany corpusFastAPI+ React
Project Brain: background workers pull the company's own content into Postgres and Supabase with Redis on the fast path, retrieval runs over that corpus with open-source models fine-tuned on top, and the answers surface through a FastAPI backend behind a React frontend.

More

Email

maxnaid41@gmail.com