Point Nemo
Home | About | Internship
August 2025
My summer 2025 internship at Leidos Dynetics in Huntsville, Alabama. Six weeks of building an LLM-based conversational pipeline, learning from incredible mentors, and diving deep into AI and machine learning.
Week 1
My first week in Huntsville was filled with meeting new people, learning new things, and seeing things I didn't even know existed. From the very start, Dynetics has created an environment that's not only technically exciting but also genuinely supportive. Everyone I've met has been welcoming, curious about our work, and more than willing to share their own experiences and advice.
My friends and I arrived on Monday (5/26), moved into the Airbnb, and immediately got to work exploring the materials Mundy had given us to study. The following few days were focused on getting oriented, meeting new people, setting up our spaces, and learning more about our projects. Mundy was great about giving us a strong start and making sure we were comfortable.
What stood out to me the most this week was the company culture. I could've never imagined how inviting and just generally nice the community at Dynetics is especially the AI and Machine Learning Branch. I'm more than excited to keep building on this momentum and continue learning new things every day.
Week 2
My second week is where I really got to dive deeper into my actual work and start getting my hands dirty. My project in particular dealt with integration into a larger code base. This is usually an incredibly daunting task; however, the engineers here were eager to help me out and explain the entire code base to me so I got a great understanding of the bigger picture. This allowed for quick turn around on my own code -- less time spent having to parse other coders stuff meant I could focus on making mine the best it could be. By the end of the week I had a nice MVP that I could show off to my mentors.
Finally, on Friday, my peers and I gave presentations to our mentors about our project, where it is, where it's going, and what we've learned. This was one of my favorite parts of the week. I got to not only present my project and get very valuable feedback on it, but I also got to watch my friends present theirs. It was really cool to see. The plan is to do this every week until the end of the internship as to prepare us for our final presentation in front of the big guys.
Week 3
I figured I'd start this week's post by diving deeper into my project. The core idea is to build a modular LLM pipeline that can integrate easily with other systems. Up until this week, my pipeline could:
This basic version worked but had several flaws and missing pieces. So this week, I focused on expanding its capabilities. I added:
Getting live data flow working meant digging into HTTP requests, port forwarding, and optimizing speed and bandwidth. I designed a batching system to send data in chunks instead of one item at a time, which made a big difference.
The imperative input feature started out simple: I used a small LLM, did some prompt engineering to pick actions from a list, and parsed the output. But by the end of the week, I overhauled it completely. Originally, adding a new action required editing multiple files and wasn't intuitive. Now, with a cleaner object-oriented approach, I created an Action class. Developers can simply extend this class to define new actions with a name, description, example, and the actual code to run. An ActionDispatcher handles registering and calling them. This is much more elegant and I'm very happy with the outcome.
The speech-to-speech pipeline was surprisingly straightforward. I used OpenAI's Whisper Small for speech-to-text and a separate text-to-speech model, then connected them a good example of how clear pipeline design makes life easier. The size of these speech to text models blew me out of the water. I mean speech to text is halfway magical and these models are in the low millions of parameters!
The biggest challenge of the week, by far, was setting up the client-side question and answer system. Without getting too deep into the weeds: I needed two servers to communicate securely, which required port forwarding and a custom proxy server to handle traffic. It took a full day of troubleshooting, overhauling, rubber ducking with Daniel, and two hot chocolates, but it finally worked.
Overall, this week really pushed me to apply and expand my knowledge across multiple areas and I'm excited to keep improving this system.
Week 4
Compared to last week, this one felt slower on the coding front but was just as valuable in other ways. I spent quite a bit of time in meetings discussing the future direction of my project. Nobody loves meetings, but I really appreciate how much thought everyone (especially Mundy) puts into making sure I have meaningful, interesting tasks and a clear path forward. It means a lot to feel supported and trusted to steer this project where I think it should go.
After a few demo sessions showing my rough-but-working pipeline, I got a lot of useful feedback and ideas for new features. I also carved out some time to read more papers and deepen my understanding of the core topics behind my project. On the coding side, progress was smaller but solid: I finalized a stable launch sequence and tinkered with the UI to make it cleaner and more intuitive.
One highlight each week is being part of Mundy's group activities: our presentations, journal club, and the AI/ML boot camp. This week's presentation session went really well. Mundy helped me polish my slides with clearer graphics and better structure. After I presented, Daniel gave a mini-talk about P vs NP -- a classic computer science problem that was followed by a fun discussion and made me look forward to my time to present even more.
For journal club, we read a paper introducing the basics of Information Theory. I'd never formally studied it before, but after reading and watching a few follow-up videos, I felt like I started seeing Information Theory concepts everywhere. It's honestly amazing how one paper can give you a fresh lens on the world. Huge shoutout to Mundy for picking such a thought-provoking read.
To wrap up the week, we had our AI/ML boot camp session, where Mundy broke down how loss functions and gradient descent work in training and backpropagation. Even though I'd seen these ideas before, getting to revisit them with clear explanations and examples really helped solidify my understanding.
All in all, this week felt like laying down the groundwork for bigger things ahead both in terms of technical progress and deepening my knowledge base. Excited to see what next week brings!
Week 5
If Week 4 was about laying groundwork, Week 5 was about discovering just how much of that groundwork I could rip up and rebuild smarter. Early in the week, I explored a larger LLM, Granite 3.3 (2.5B parameters), and followed a LangChain tutorial on agentic RAG. It was decent, not production-level, but enough to make me realize something slightly embarrassing: I'd been accidentally recreating half of LangChain over the past few weeks. On the bright side, it meant I now understood the pieces deeply enough to trim away a ton of complexity.
That realization hit me one night during dinner and I ended up locking in super hard for like 5 hours and completetly rebuilding my pipeline around a much simpler approach: using a vector database and a k-NN-based encoder to handle tool selection and slot filling. The result is something surprisingly powerful, something like a lightweight Siri-for-missions. It's faster, leaner, and often outperforms the full LLM-based approach in practical scenarios. Huge win.
On the project coordination side, I had yet another direction-setting meeting, but this one felt actually really ended up with some useful infromation. I shifted focus to integrating my chatbot into ADAPT GO and got it running with real functionality, including generating and editing human constraint objects on the fly.
I also cleaned up the UI and polished the interaction between ADAPT and my NLP model. I now have a personal GitLab repo where I'm tracking the project and keeping things modular. One fun breakthrough was getting the bot to handle incomplete user instructions: like when someone says "add a task," and it follows up naturally to fill in the missing details. It's a small thing, but it really makes the system feel conversational.
There's still room for refinement. The slot filler struggles with subtle semantics like confusing "set Raptor 2's priority to 5" and interpreting it as priority 2. I'm experimenting with regex and semantic cues for now, but I may later explore training a more robust model with LLM-generated synthetic data. I'm also starting to think about Chain-of-Thought-style reasoning and how to incorporate pre-planning logic without making the system too slow or brittle.
Overall, this was my most transformative week yet. I tore down and rebuilt the pipeline, implemented key ADAPT integrations, and came away with a system that's already looking solid for demos. There's a long road ahead, but I'm starting to see how this all fits together and I'm genuinely excited to keep pushing.
Week 6
This week started off rough. I hit a wall of uncertainty... not technical, but directional. The tasks I had lined up (confidence thresholds, parallel instructions) felt more like maintenance than innovation, and I wasn't excited to work on them. I kept circling the same question: what exactly should it do? The answer didn't come quickly, and it left me feeling pretty stuck.
I filled some of that time doing light work: tweaking slides, reading papers, adding a feature here and there (like handling low-confidence queries gracefully). But it wasn't until later in the week that things finally started clicking again. I cleaned up the tool system, refactored big parts of the pipeline, and got the project running locally with database integration (a surprisingly satisfying milestone). Presenting to Andrew helped too, even if his reaction was more politely intrigued than wowed.
One key turning point was running into major issues when using actual entity IDs. The model began hallucinating parameters and tools, and nothing behaved the way I expected. That frustration drove me to rebuild AGAIN. And through some more research, I learned about zero-shot intent classifiers and realized that was exactly what I needed. I got one running and, for once, it just worked. It wasn't perfect, but it worked out of the box. So, I built fallback mechanisms and eventually landed on a really cool, new architecture.
Instead of relying on a single model with confidence thresholds, the new pipeline pools together multiple intent-scoring strategies: zero-shot classification, keyword matching, semantic similarity, spaCy pattern matching, and more. I score each method individually, weight them, and then rank to find the best match. It's ensemble intent classification and it's shockingly effective. The system is now both more accurate and more resilient. It's also way more satisfying to use.
Despite the slow start and occasional "why am I doing this" moments, this ended up being one of the most productive weeks yet. I still have bugs to squash and weird behaviors to figure out, but the pipeline feels smarter, cleaner, and closer to something I'd be proud to demo. That's a good place to be heading into the final stretch.
© 2026 Parker Hance