This year will be on rebuilding and growth. After a challenging 2025, I am looking forward to a fresh start in 2026. After living in 4 separate cultures - Singaporean, American, Australian and Dutch, I have a privileged perspective on cultures, lives, countries.
Granted they are all a snapshot in time, formed through my own personal lens and limited by my sphere of comprehension at that time, I do feel I have a broader perspective than most. I am exceedingly thankful for the life I have led so far, I have taken on a lot, challenged a lot and overcome a lot. Moving forward, I want to think about which direction I want to take the next 5 yrs.

Chatbot Deployment Update: 15 Jan tweaks

I spent a few days rounding back to this and thinking on how I can improve it. I am still learning about AI but I feel that the chatbot gives limited examples of my work/portfolio, and also doesn’t do enough filtering of irrelevant questions and naturally accepts or rejects what questions it can answer, what it is trained on. The current problems:

• configuration issues (k=5 is too low, no chunking)
• filter design (too permissive for personal questions when there is little personal data in dataset)
• prompt emgineering (summarization towards personal skills)
• lack of testing (no established baseline)
• multiple LLMs (anthropic never trains on api data, openai stopped since march 2023 and ollama is 100% private)

Currently, the costs for the current set up is really low (was my goal). 79 mdx files, about 600kb of text, 6818 lines = 145k tokens at 580kb / 4 chars per tokens It fits well in a single LLM conext window for Claude 2.5 Sonnet ad Haiku, will need chunking for gpt-4o-mini. Cost per query is $0.00077, assuming 1k querues: $0.77 per month.

If I ran this through as a document to an LLM (much alike notebook LM): Without Prompt Caching:

• GPT-4o-mini: $0.022/query → $22/month (28x more)
• GPT-4o: $0.364/query → $364/month (473x more)
• Claude Haiku: $0.117/query → $117/month (152x more)
• Claude Sonnet: $0.437/query → $437/month (567x more)

With Prompt Caching (Anthropic 90% discount, OpenAI 50% discount):

• GPT-4o-mini: $0.011/query → $11/month (14x more)
• Claude Haiku: $0.012/query → $12/month (15x more)
• Claude Sonnet: $0.044/query → $44/month (57x more)

Accuracy Analysis

Would it be more accurate? YES, significantly:

✅ Zero retrieval failures - Every document always available ✅ Better “list all” queries - LLM sees everything, can’t miss projects ✅ Cross-document reasoning - Can compare/synthesize across all content ✅ No chunking artifacts - Full context, no split information ✅ No scoring bias - No keyword ranking affecting visibility

Potential downsides:

⚠️ Lost in the middle - With 145k tokens, some attention dilution ⚠️ Latency - Processing 145k tokens slower than 5k tokens (~2-4 seconds vs less than 1 second) ⚠️ Scalability limit - Won’t work if content grows to 500k+ tokens ⚠️ Cache invalidation - New content requires full re-cache

If i want a hybrid approach with the two:

• 20% of queries are “list all” type: 200 * $0.012 = $2.40 (“What are your list of …”, “What are all your… ” questions)
• 80% of queries are specific: 800 * $0.00077 = $0.62
• Total: ~$3/month (4x current cost, but solves both problems)

The flow chart would be a multi-agent system that looks like this:

Original Multi-Agent Design

Optimized Multi-Agent (Cheap)

Savings: $5.60/month (70% reduction!)

How Behavioral Analysis Works

You asked: “how to draw connections and analysis about a person through blogs”

Strategy:

1. Load blog posts (recent 20 posts, ~60k tokens)
2. Extract themes (first pass with Claude):

• Leadership examples
• Problem-solving approaches
• Values and motivations
• Lessons learned

3. Synthesize answer (second pass):

• Use STAR format (Situation, Task, Action, Result)
• Base on actual blog content
• Show self-awareness and growth

Example: Q: “How do you handle technical disagreements?”

Behavioral Agent: → Scans blogs for conflict resolution stories → Finds: “In my post about ManaBurn architecture decisions…” → Extracts pattern: “I prefer data-driven discussions + prototypes” → Synthesizes: “I approach technical disagreements by [STAR response based on actual blog examples]”

Call me cheap but $8/mth is more than I want to pay. Reiterating. More tomorrow.

Personal Goals for 2026

Physical

Getting into peak form in terms of strength, weight-bearing workouts, flexibility and cardio.

Mental

Onboarding a meditation practice, journaling practice, and reading habit.
Finding a comfortable groove that allows for reflection and growth.

Professional

Looking for an evironment that challenges me, focuses on healthy growth.