Hi, I’m Jackson.

I’m a CS junior at the University of Florida, interning at Datadog starting in January. I build backend and data systems, and I am currently doing research on data collection methods for training language models.

See my projects GitHub ↗ LinkedIn ↗
01 /

Experience

INCOMING · JAN TO APR 2027 · NEW YORK, NY

Datadog

Software Engineer Intern
Incoming winter, team to be determined.
MAY TO JUL 2026 · MELBOURNE, FL

Northrop Grumman

Software Engineer Intern
  • Built C++ clients for a distributed pub/sub system, using incoming and outgoing queues around a central platform loop that updates state from messages.
  • Replaced individual client classes with one reusable template, instantiated per client with its own message type, registration config, and optional client-specific functions.
MESSAGE FLOW
NetworkReceptionistPlatformPublisherINCOMINGtyped messagesOUTGOINGtyped messagesupdates statelog (optional) NetworkReceptionistPlatformPublisherincomingtyped messagesoutgoingtyped messagesupdates statelog (optional)
JUN 2025 TO MAY 2026 · NAPLES, FL

Naples Comprehensive Health

Software Engineer Intern
  • Built a web tool that merges care gap spreadsheets from 8 insurer formats into one deduplicated master list using saved per-insurer configs, cutting duplicate audit work by about 20%.
  • Moved heavy file processing from Flask to AWS, cutting spreadsheet merge time from 60 to 4 seconds with Lambda and running batch PDF sorting through S3 + Lambda.
MERGE FLOW
invokereturnedInsurer AInsurer BInsurer CFlaskserverAWS Lambdasimplifiedfor sheet in sheets:rows = remap(sheet, config)master = concat(master, rows)master.drop_duplicates()Master listdeduplicated invokereturnedInsurer AInsurer BInsurer CFlask serverAWS Lambdasimplifiedfor sheet in sheets:rows = remap(sheet, config)master = concat(master, rows)master.drop_duplicates()Master listdeduplicated
02 /

Projects

01 / In progress

Text2SQL

A from-scratch 1B-parameter LM trained on a mixture of SQL and English, powering an autocomplete and text-to-query wrapper.
WHYInterested in LLMs end to end. A narrow use case makes a small model more useful than a general one. NOWFine-tuned Laya classifier that scores web pages on SQL quality. Pre-training corpus at roughly 150M pages and growing. NEXTMore training data, including pure code, then a training framework.
Work in Progress Code
PREVIEW · SAMPLE DATAbikes, brands
ENGLISH TO SQL
db.query(text2sql(
"lightest full suspension bikes
with 150mm+ travel"))
SELECT name, weight_lbs FROM bikes
WHERE travel_mm >= 150
ORDER BY weight_lbs LIMIT 5;
nameweight_lbs
Hightower30.8
Stumpjumper EVO31.2
SB15031.4
Slash32.1
Fuel EX 833.0
AUTOCOMPLETETab
SELECT b.name, br.name
FROM bikes b JOIN brands br ON
br.id = b.brand_id
02 / Data pipeline

AI Research

Does deduplicating LLM training data quietly erase specialized text? An audit of 99.4M Common Crawl pages.
PROBLEMLLM pipelines use one fuzzy-dedup threshold for all text, never tested by domain. RESULTHypothesis not supported. Dedup changed legal’s topic diversity within 1 percentage point of general’s at every threshold. NEXTTrain two ~1B LLMs on data deduped at 0.6 and 0.9, then test how they perform.
99.4M
pages after filtering
50,104
legal pages found
~55 h
one HPC run
FLOW
Source & filter
26,776 files, 99.4M pages
Classify legal
TF-IDF screen, then 421M model
MinHash dedup
4 thresholds, 0.6 to 0.9
Topics & entropy
BERTopic vs random removal
the stage that does the heavy lifting
03 / Database

JackyDB

A column-based storage engine and SQL query executor written in Go.
BUILTA custom binary file format with auto-typed columns, loaded straight from CSV. RESULTNarrowest-fit integer and float types shrink files. Larger buffers sped up ingestion. NEXTRow groups with a footer layout, then SQL execution and a DuckDB benchmark.
42%
avg compression vs CSV
29.5x
faster ingestion vs no buffering
.JDB FILE LAYOUT
magic, version
row grouprepeats
NUMERIC COLUMNS
null bitmapraw values
STRING COLUMNS
null bitmapoffset mapblob
group footer
has nullstightest typeblob length
global footercolumn names, final types, group offsets
footer length, magic
04 / Systems

Game Boy Emulator

A cycle-accurate Game Boy emulator written in C++.
BUILTFull CPU instruction set, scanline-based graphics, and cartridge memory banking. RESULTRuns real ROMs with keyboard controls and save files.
500
opcodes implemented
21x
faster than original hardware
DATA FLOWONE INSTRUCTION
MEMORY
Interrupt check
Handles button presses and other urgent events first.
Fetch and decode
Reads an instruction and finds its code.
Execute
Math, data manipulations, jumps.
Graphics
Draws one screen line at a time.
Timers
Counts time for clocks and countdowns.
pending events
instruction
reads, writes
reads graphics
writes status
updates clock
next instruction
03 /

About me

OUTDOORS

Mountain biking

My Trek Fuel EX 8 on a limestone trail

Outside of tech, I mountain bike. I love being outside and going fast, and this lets me do both at once. I started on my dad’s old Gary Fisher, and as I got more into it I upgraded to a Trek Fuel EX 8. It’s my break from school and technical work, and one more thing to practice and get better at.

WHERE I RIDE
Markham Park
Alafia River State Park
MUSIC

My picks

I’ll listen to any sub-genre of rap, with Yeat by far my favorite. Also topping my list are Young Thug, Kanye, Che, OsamaSon, prettifun, Ken Carson, and a lot more. I listen more for the beat than the lyrics, since good production is what makes a song stand out to me. Earlier this year I saw both Kanye and Yeat live.

Up 2 Më album cover
Up 2 Më
Yeat
2093 album cover
2093
Yeat
Slime Season 3 album cover
Slime Season 3
Young Thug
Psykotic album cover
Psykotic
OsamaSon
More Chaos album cover
More Chaos
Ken Carson